<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Batuhan Sener on Medium]]></title>
        <description><![CDATA[Stories by Batuhan Sener on Medium]]></description>
        <link>https://medium.com/@batuhansenerr?source=rss-2256a8221220------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*jbsCYJ9u8Cn1YW33AQAsKg.jpeg</url>
            <title>Stories by Batuhan Sener on Medium</title>
            <link>https://medium.com/@batuhansenerr?source=rss-2256a8221220------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 13:30:22 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@batuhansenerr/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Inside DeepSeek’s Super-agent: One Router, Many Workers, No Sessions]]></title>
            <link>https://medium.com/@batuhansenerr/inside-deepseeks-super-agent-one-router-many-workers-no-sessions-da03818d4540?source=rss-2256a8221220------2</link>
            <guid isPermaLink="false">https://medium.com/p/da03818d4540</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Batuhan Sener]]></dc:creator>
            <pubDate>Tue, 03 Nov 2026 10:47:34 GMT</pubDate>
            <atom:updated>2026-11-03T10:47:34.432Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qin4u0gw3HRNt6CIuPt34A.png" /></figure><p><em>Plus: Gemini 4 Carbon tops Google’s own table, NVIDIA opens a 40-language speech model, and a TPU answers from orbit.</em></p><p>DeepSeek’s Harness team has published a detailed design for an agent that runs all day across chat, GitHub and email, and they back it with their own experiments. Most of this issue walks through it: what the design does, why they built it this way and how each part works. The issue covers the three weeks from October 19, and everything in it was posted between October 28 and November 1.</p><p><strong>In this issue:</strong> 2 top stories · 2 models · 1 tool · 1 quote</p><h3>Top Stories</h3><p><strong>DeepSeek’s blueprint for an agent that outlives the chat session</strong></p><p>On November 1 the DeepSeek Harness team published “Exploring Super-agent Design: Towards a More Autonomous Agent System”. It runs as an X Article, machine-translated from Chinese and written by Yujia Qiao. The team says up front that it covers architectural ideas, not existing DeepSeek Harness features. It is still a concrete public blueprint for an agent that lives inside a team’s chat and GitHub instead of a terminal tab, so it gets most of this issue.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*1U2QLoVWKiEXayWN" /></figure><p><strong>The problem: human attention</strong></p><p>Today’s coding agents treat the session as the unit of work, and sessions share nothing. If you run several tasks in parallel, you switch between them by hand and remember what each one knows. DeepSeek puts the limit at 3 to 5 active sessions a person can follow at once. In their words, a single task is limited by the model, but multitasking is limited by human attention. Adding CLIs, MCP or connectors to a local single-session agent doesn’t fix this, they argue. The session still blocks on long operations, contexts from different tasks get mixed together, and everything stops when the laptop goes offline.</p><p><strong>The goal: one agent with four properties</strong></p><p>Their answer is a single long-running agent that keeps its context across sessions and channels and works on several tasks at once. They call it a Super-agent and give it four properties:</p><ul><li><strong>Continuous:</strong> it feels like one agent wherever you talk to it. Ask for Plan A in chat, switch to Plan B in a pull request comment, and it adjusts.</li><li><strong>Productive:</strong> it runs several complex tasks in parallel and takes the next instruction before the last one finishes.</li><li><strong>Proactive:</strong> it watches events in its environment, acts on what is worth handling and stays quiet otherwise.</li><li><strong>Persistent:</strong> it is always online, connected to real work environments, and recovers from failures by itself.</li></ul><p>No single ReAct loop delivers all four, so the rest of the essay is about the system around the loop. Most of it borrows classic ideas from computer systems design.</p><p><strong>How it works, module 1: a Router and many Workers</strong></p><p>A long-lived Router session handles every human conversation and dispatches work. It never executes tasks itself and can run without a Bash tool, so DeepSeek finds a lighter model is usually enough. Workers are isolated sessions, one per task, with a full tool set and the strongest model available. Workers never talk to people. They report to the Router, and the Router writes every reply. The team separates this from an “Agent Team”: the Router doesn’t break a task into steps, it only dispatches tasks and tracks how they relate. Sessions turn into an implementation detail the user never manages.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*IQzdRtMzyi3vVnFL" /></figure><p><strong>Module 2: events instead of prompts</strong></p><p>The agent’s main input stops being prompts typed by a person. Chat messages, emails, calendar reminders, review comments and CI results arrive as “Triggers”. A code filter (the Fast Path) drops most of the noise by source, group and allowlist before any model runs. An LLM (the Slow Path) judges the rest against policies written as prompts, which people can switch on or off in plain language. Consecutive events on one pull request, chat topic or email chain form a Thread. The Worker created for a Thread subscribes to it, and code passes follow-up events straight to that Worker without touching the Router.</p><p>Context gets treated like a cache. “Inline” puts what the model will probably need soon into the window; “Spill” leaves the rest outside behind a pointer the model can follow. Because every object carries the IDs of related objects, the Router can walk back to an old task hop by hop. DeepSeek calls this Chain of Recall. The agent also acts ahead: their internal user group runs a “silently reproduce and fix” policy. When someone reports an issue, the agent tries to reproduce it in the background and says something only after it has confirmed the issue and found a simple fix.</p><p><strong>Module 3: talking while working</strong></p><p>In a classic ReAct loop, a run() tool call blocks until the tool returns, and the agent can&#39;t answer anyone in the meantime. DeepSeek replaces long calls with asynchronous start() and stop(). start() returns &quot;Started&quot; right away, the final result arrives later as a new event, and the agent keeps answering people in between. DeepSeek Harness already implements this, but it is not on by default. Each channel also gets its own norms: short in chat, where a hard cap of 100 words per reply is a rule they say they verified, and structured, dense output on GitHub and email.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*lyS3FlM-aGKIpJT3" /></figure><p><strong>Module 4: compute, state and environment, separated</strong></p><p>The agent loop runs in stateless cloud containers and session state lives in a database, so a sleeping laptop no longer stops the agent. Execution environments stay where they are. Each machine runs a lightweight daemon that keeps no session state and only executes commands and returns results. That lets the agent work on your own machine, uncommitted changes and local credentials included. They call it BYOM, Bring Your Own Machine. When an intermittently online device reconnects, the cloud agent runs the tasks that need it. DeepSeek credits Anthropic’s “Decoupling the brain from the hands” with the same approach for managed sandboxes.</p><p><strong>Module 5: a Watchdog outside the system</strong></p><p>A separate session, driven by scheduled messages, handles infrastructure failures such as API errors. It also asks the model whether each Worker is making real progress or has drifted from the goal, because many agent failures aren’t crashes: the process runs fine while the model loops. The Watchdog is itself an agent with tools, so maintainers can ask it in plain language why something failed.</p><p><strong>What DeepSeek measured</strong></p><p>The first experiment fed four designs the same stream of about 300 group chat and GitHub CI messages, about 20 of them real work requests, 5 runs each. A tool that only consumes time stood in for coding, so the test measures scheduling, not coding skill. Super-agent finished tasks in 12.4 minutes at p50 and 15.7 at p95. A single session with a subagent tool and a neutral prompt took 20.4 and 29.5, and it didn’t call a subagent in any of the 5 runs. With a prompt that pushed delegation it took 13.8 and 29.0. It delegated in 4 of 5 runs; in the fifth it judged the work “small” and did it itself. Super-agent also used 23.2M cache-read tokens per run, against 6.1M and 10.8M for the single-session designs. Without pass-through, speed stayed about the same, but the Router made about 30% more decisions and tokens rose to 31.8M.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1009/0*3m05nO8mAotXg8wf" /></figure><p>The second test stressed the Router alone. As triggers rose from about 1 to about 40 per minute, events that piled up during one inference turn were merged into the next call. Average batch size grew from 1.0 to 2.8 while time per request only grew from 2.5 s to 3.9 s. Time per decision fell from 6.4 s to 1.6 s. At DeepSeek’s team size, one Router session is enough.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*2vryNSFUmiEp9jfc" /></figure><p><strong>What it means</strong></p><p>I’d build delegation into the structure instead of asking for it in a prompt. The neutral-prompt session that never handed off work once in 5 runs is the result that should worry anyone running a single agent with a subagent tool. You can copy the Router/Worker split and the Fast Path filter without adopting DeepSeek’s stack.</p><p>The team lists what is still open: long-term memory over months of operation, what interface replaces the chat window, how agents with different permissions talk to each other, and security. If a server intrusion or a prompt injection compromises the cloud agent loop, every endpoint device connected to it could be taken over. Credentials on an endpoint can also leak through the model’s context during normal execution, without any breach.</p><p><a href="https://x.com/DeepSeekHarness/status/2116857257245987236">Source →</a></p><p><strong>Google ran a prompt on a TPU in orbit and got a haiku back</strong></p><p>Sundar Pichai posted a “hello world” milestone for Project Suncatcher on October 30. A few weeks after launching a prototype satellite, Google sent a command up to orbit, booted the TPU, ran a prompt and received the first complete response back on Earth: a haiku. The post comes with a 16-second video and no technical detail. For now it proves the full loop works: uplink, boot, inference, downlink. It says nothing yet about whether compute in orbit can compete with a data center on the ground on cost, power or latency. Treat it as a milestone to track, and wait for the numbers before it changes any plan.</p><p><a href="https://x.com/sundarpichai/status/2116214866444943853">Source →</a></p><h3>New Models</h3><p><strong>Gemini 4 Carbon: Google’s new top score on the Artificial Analysis index, two weeks after Argon</strong></p><p>Logan Kilpatrick announced Gemini 4 Carbon on October 28 as “a big jump over Argon which we released just 2 weeks ago”. Google’s table puts it at 59 on AA Index v5.0, against 58 for Claude Opus 5.5, 57 for GPT-6 Astra and 50 for Argon. The biggest gain in the table is long context: 96.2% on GraphWalks, against 84.2% for Argon. The lead on the headline index is one point, so two flagship releases two weeks apart are a reason to re-run your own evals, not to switch on the spot.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*f65dpElpAW7smiOH" /></figure><p><a href="https://x.com/OfficialLoganK/status/2115489284404940898">Source →</a></p><p><strong>Nemotron 3.5 Transcribe: one open speech model for 40 languages and hour-long recordings</strong></p><p>NVIDIA released it on Hugging Face on October 28 under OpenMDW-1.1. It is a hybrid Mamba-Transformer mixture-of-experts model with about 4.3B active parameters. One model handles transcription, translation, speaker-attributed transcripts, timestamps and audio Q&amp;A, chosen per request by instruction. On the nine English Open ASR Leaderboard sets, NVIDIA reports a mean WER of 4.71 against 4.94 for Qwen3-ASR-1.7B. If you chain separate ASR, diarization and translation steps for meetings, this is the model to test for collapsing them. It is offline only, so it won’t do live captions.</p><p><em>Heads up: NVIDIA measured every number, on October 12, before release. The comparison model has 1.7B parameters against Nemotron’s 4.3B active. The card’s name says 30B-A4B while its spec table says 34B total, audio encoder included. The serving example targets one H100 80 GB.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Uz9JhkqqADvDDG9I.png" /></figure><p><a href="https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B">Source →</a></p><h3>Tools &amp; Libraries</h3><p><strong>dsh-superagent-prototype</strong>: DeepSeek’s MIT-licensed research prototype of a proactive coding agent for one GitHub repository, packaged for DeepSeek Harness.</p><pre>Read SETUP.md in this repository and set up dsh-superagent with me.</pre><p>The README asks you to paste that line into a coding agent instead of running install commands. The agent listens to GitHub and optionally Feishu, opens work items for allowed requests, works in its own session and directory, and replies in the original conversation, for example with a pull request. A web panel shows items that need a person. DeepSeek says it is not an official product, not production-ready, adapted from the internal version, and that coding agents did much of the engineering. Harness itself starts with npx @deepseek-ai/dsh web. Run it on a throwaway repo first; the security warnings in the essay apply here too.</p><p><em>Who it’s for: engineers who want to study an event-driven agent on a sandbox repository before building their own.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*vYWae8o_h6iWmNcw" /></figure><p><a href="https://github.com/deepseek-ai/dsh-superagent-prototype">Link →</a></p><h3>Quote of the Week</h3><blockquote><em>“Sessions should be an implementation detail of the system, not a mental burden on its users.”<br> — DeepSeek Harness team, “Exploring Super-agent Design”</em></blockquote><h3>What To Do This Week</h3><p><strong>Try:</strong> test DeepSeek’s Router/Worker split on your own repository in Claude Code.</p><ol><li>Pick three small, independent tasks in one repo (a bug fix, a missing test, a doc update) and write down what “done” means for each.</li><li>Run them one after another in a single Claude Code session. Note the wall-clock time and how many times you had to step in.</li><li>Start a fresh session and tell Claude to act only as a dispatcher: hand each task to its own subagent, run long commands such as the test suite in the background, and report back only results and questions.</li><li>Compare time, interruptions and token use. Expect higher token use: in DeepSeek’s experiment the split cost about 2 to 4 times the cache-read tokens.</li><li>Keep the dispatcher setup if it saved your attention, and drop it if the tasks were too small to benefit.</li></ol><ul><li><strong>Read:</strong> <a href="https://x.com/DeepSeekHarness/status/2116857257245987236">Exploring Super-agent Design: Towards a More Autonomous Agent System</a></li><li><strong>Watch:</strong> <a href="https://www.youtube.com/watch?v=hoaoQxQCoi8">Let’s have some fun !!</a></li></ul><p>If you would like to consult or get in touch, you can reach me via <a href="https://www.linkedin.com/in/yigitbatuhansener/"><strong>LinkedIn </strong></a>or <a href="http://ybatu.sener@gmail.com/"><strong>email</strong></a>.</p><p>If you enjoy my content, please consider showing your support by clapping for my articles or buying me a coffee. Your support is greatly appreciated.</p><figure><a href="https://buymeacoffee.com/batuhansenerr"><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Nh1owl6a3HdaRB_Y0y04pw.png" /></a></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=da03818d4540" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Briefing, Oct 19–25: Fewer Agents, More Measurement]]></title>
            <link>https://medium.com/@batuhansenerr/ai-briefing-oct-19-25-fewer-agents-more-measurement-55e7a5616198?source=rss-2256a8221220------2</link>
            <guid isPermaLink="false">https://medium.com/p/55e7a5616198</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Batuhan Sener]]></dc:creator>
            <pubDate>Sat, 24 Oct 2026 20:33:16 GMT</pubDate>
            <atom:updated>2026-10-24T20:33:16.350Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ORp49e8iDHwXAdufhmVNmw.png" /><figcaption><em>Anthropic argues for fewer subagents, DeepSeek open-sources its inference stack, and three OCR releases compete on documents.</em></figcaption></figure><p>This issue covers October 5 to 25, and nearly every item in it was posted between October 15 and 23. Anthropic published two engineering pieces that give the same advice from different angles. Both say to take layers out of your agent setup and keep only the changes your numbers support. For teams parsing documents, three OCR releases in this issue compete on very different hardware budgets.</p><p><strong>In this issue:</strong> 2 top stories · 3 models · 4 tools · 5 quick hits · 1 quote</p><h3>Top Stories</h3><p><strong>Claude Code’s team moved most code reading to machines, and its revert share fell</strong></p><p>Claude writes about 90 percent of Claude Code’s own code, and from June to September merged pull requests more than doubled while the share that had to be reverted fell by more than 40 percent. Claude reviews the diff in a fresh-context subagent before a PR opens. Every PR then gets an AI review whose findings other agents verify, plus a separate security pass. A classifier sets the risk tier, and with it the number of required human approvals: none, one or two. After the team changed how reviews converge, PRs needing five or more rounds fell from 8.9% to 1.5%. If you already run an AI reviewer, track human time per change the way this team does, with revert and incident rates as the guardrail.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*yzzjCcq4yuE2ZENk.png" /></figure><p><a href="https://claude.dev/blog/solving-code-review-at-anthropic/">Source →</a></p><p><strong>Anthropic tells enterprises to fold their subagents back into one loop</strong></p><p>Anthropic’s Diane Sarkis writes that in enterprise comparisons, a single agent with skills and tools often beat both the one-big-prompt design and the subagent design, at lower cost and latency. One customer measured a roughly 2x faster runtime after moving from its multi-agent harness to a single loop. The post’s warning sign is an architecture that copies the org chart, with one subagent per team. The migration path it lays out starts with an end-to-end eval (10 golden examples is enough to start). Next you expose subagents as tools, then fold them in one at a time as skills and MCP tools and compare evals after each step. Subagents still make sense when context bloat measurably hurts the loop, for parallel work with disposable context, or when instructions conflict.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*jHmlp1zL-AkSWQ9r.png" /></figure><p><a href="https://claude.com/resources/articles/building-agents-when-to-fold-subagents-into-your-main-agent-loop">Source →</a></p><h3>New Models</h3><p><strong>DeepSeek-V4.1-Flash-1017: a minor update with a big jump on Terminal-Bench</strong></p><p>DeepSeek uploaded V4.1-Flash-1017 to Hugging Face on October 17 under the MIT license: 763B parameters and a 1M-token context. On Terminal-Bench 4.0, run in the Claude Code harness, it scores 42.3 against 31.2 for the previous V4.1-Flash and 39.9 for GPT-5.6 Sol. Opus 5 still leads at 51.8. A “minor update” that gains 11 points on terminal tasks deserves a rerun on your own agent evals before you write it off as a patch.</p><p><a href="https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017">Source →</a></p><p><strong>MiniMax M3.1 Flash: a third of M3’s cost, at $0.15 per million input tokens</strong></p><p>MiniMax released M3.1 Flash on October 19 and says it runs at 1/3 the cost of M3 and 3x the speed. API pricing is $0.15 input and $0.60 output per million tokens, with cache reads at $0.015. It scores 62.2 on Automation-Bench and 76.2 on BankerToolBench, tied with Opus 5 there. MiniMax says Space Bunny, a model that topped leaderboards under that name, was an earlier version of it. The weights are now on Hugging Face under the minimax-community license. For high-volume agent work, the price makes it worth a test.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*SBihUoyZWDHStSgk" /></figure><p><a href="https://x.com/MiniMax_AI/status/2112150867868541221">Source →</a></p><p><strong>d3-turbo: a decision model that answers 255 questions about one image in 2.12 seconds</strong></p><p>The vLLM Semantic Router team released d3-turbo, a 27B multimodal model under Apache-2.0. It writes no text. It reads the input once and returns a probability for each answer, for routing, yes/no, scoring and bounding-box questions. On one AMD MI325X it answers 255 questions about one photo in 2.12 s, against 32.94 s for d3-flash, which reads the image once per question. If your pipeline sends an LLM many small classification calls about the same input, this is the pattern to borrow.</p><p><em>Heads up: The speed chart is the team’s own. With a single question d3-turbo is slower, 302 ms against 154 ms, and the 70.2 text score on the Jev Decision Index is the team’s own estimate, not an official board result.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*12rIhx5q3pAQlZ0b.png" /></figure><p><a href="https://x.com/XunzhuoLiu/status/2112343603209261092">Source →</a></p><h3>Tools &amp; Libraries</h3><p><strong>DeepInfer</strong>: DeepSeek’s in-house inference framework, which serves DeepSeek-V4.1-Flash in production, now open source under Apache-2.0.</p><pre>export DEEPINFER_ACCELERATOR=nvidia &amp;&amp; ./build.sh</pre><p><em>Who it’s for: teams self-hosting DeepSeek-V4.1-Flash on NVIDIA SM100 GPUs or Huawei Ascend 950 NPUs. It runs no other model, and this release leaves out the KV cache system and prefill/decode disaggregation, so treat it as a reference for DeepSeek’s serving tricks more than a drop-in vLLM replacement.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YIizHXfSwCeqalhm.png" /></figure><p><a href="https://github.com/deepseek-ai/DeepInfer">Link →</a></p><p><strong>Indra OCR 1.1</strong>: Datalab’s 1M-parameter OCR model, which runs on CPU through ONNX Runtime and reads 80+ languages and handwriting.</p><pre>pip install indra-ocr</pre><p><em>Who it’s for: anyone running OCR at volume without a GPU. Datalab measured 43.9 pages per second on 32 CPU cores, against 2.9 for Tesseract 5, on its own benchmark. The weights are free for research, personal use and startups under $20M in funding or revenue; other commercial use needs a license.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*qpf3U6fBzml7Zs_h.png" /></figure><p><a href="https://github.com/datalab-to/indra">Link →</a></p><p><strong>OXR-1.0</strong>: SoMark’s open document parser, which turns PDFs and images into Markdown or JSON with layout coordinates.</p><p>pip install somark, then somark --base-url https://free-oxr.somark.ai/api/v1 --api-key dummy parse document.pdf</p><p><em>Who it’s for: teams that want to self-host a parser on a 24 GB GPU, with a free hosted API to try it first. SoMark reports 87.84 on olmOCR-bench at 558 tokens per second on one RTX Pro 6000, in its own tests. The license is Apache 2.0 with extra terms, so read it before you ship.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*N4dnDFxqaRDtyrTG.png" /></figure><p><a href="https://huggingface.co/SoMarkAI/OXR-1.0">Link →</a></p><p><strong>OCRia-1</strong>: LlamaIndex’s 4B document parsing model, now behind LlamaParse at the same price. LlamaIndex reports 87.4 on its ParseBench against 75.3 for LightOnOCR-3–4B, but on olmOCR-bench, which it did not build, the lead over olmOCR-2 shrinks to 83.5 against 82.3.</p><p><em>Who it’s for: LlamaParse users with financial tables and charts. The post doesn’t mention downloadable weights, so this is a hosted upgrade, not a model you can run.</em></p><p><a href="https://www.llamaindex.ai/blog/parse-ocria-1">Link →</a></p><h3>Quick Hits</h3><ul><li><strong>Claude Duet:</strong> Anthropic opened a waitlist research preview of a proactive agent, built around Opus 5.5, for US Pro and Max subscribers; it works in its own virtual machine and asks before it sends, pays, signs or deletes. <a href="https://x.com/claudeai/status/2112984664256241837">source</a></li><li><strong>OpenAI on COBOL and Fortran:</strong> OpenAI says its latest models are better at building and migrating software in low-resource languages such as COBOL, Fortran and Visual Basic, and asks enterprise teams which languages to tackle next. <a href="https://x.com/OpenAI/status/2112310345863930168">source</a></li><li><strong>Mercury 3:</strong> Inception’s diffusion LLM runs at 1,068 tokens per second on NVIDIA GPUs with a 1M-token context; Inception claims it beats GPT-6 Luna and Claude Haiku 5.5, but its launch post names no independent evaluator. <a href="https://x.com/StefanoErmon/status/2112222911218291051">source</a></li><li><strong>FLUX 3 Max:</strong> fal’s post-trained FLUX 3 ranks #2 on Artificial Analysis’ image-to-video board, statistically tied for first, at $10.20 per minute of 720p video against $17.40 for the base FLUX 3. <a href="https://x.com/ArtificialAnlys/status/2112266414593827110">source</a></li><li><strong>Seedance 2.5 4K:</strong> ByteDance added 4K output with 10-bit color, plus a Draft mode for cheap 480p previews you upscale only once the shot works. <a href="https://x.com/testingcatalog/status/2113667822132027903">source</a></li></ul><h3>Quote of the Week</h3><blockquote><em>“Revert and incident rates are the guardrail: if they rise, we haven’t saved human time, we’ve only moved it.”<br> — Addy Osmani, Morgan Westlee Lunt and Qing Wang, on how the Claude Code team does code review</em></blockquote><h3>What To Do This Week</h3><p><strong>Try:</strong> measure your own AI code review loop, then change one thing and keep it only if the numbers hold.</p><ol><li>Pick three to five recent pull requests in one repo. Define done as: tests green and no open finding marked important.</li><li>Run /code-review on each diff in Claude Code. Record the number of findings, how many were real bugs, how many review-and-fix rounds it took to reach done, and your minutes spent reading.</li><li>Compare the rounds: for each finding after the first round, check with git blame whether the flagged code was already there at the first review. On the Claude Code team, about 59% of what the reviewer kept flagging was code that had been there since its first pass.</li><li>Apply one change from the Claude Code post: add a REVIEW.md with your three most repeated review comments, and cap reviews at two rounds, re-reviewing only the new diff after the first.</li><li>Re-run the same PRs plus the next few new ones. Keep the change if rounds and reading minutes drop and no more reverts appear. Otherwise drop it.</li></ol><ul><li><strong>Read:</strong> <a href="https://claude.dev/blog/solving-code-review-at-anthropic/">How the Claude Code team does code review</a>, for the full adoption checklist: risk tiers, merge queue, feature flags and an eval set built from your own reverts.</li><li><strong>Watch:</strong> <a href="https://www.youtube.com/watch?v=GfOW2ItMcXE">Let’s have some fun !!</a></li></ul><p>If you would like to consult or get in touch, you can reach me via <a href="https://www.linkedin.com/in/yigitbatuhansener/"><strong>LinkedIn </strong></a>or <a href="http://ybatu.sener@gmail.com/"><strong>email</strong></a>.</p><p>If you enjoy my content, please consider showing your support by clapping for my articles or buying me a coffee. Your support is greatly appreciated.</p><figure><a href="https://buymeacoffee.com/batuhansenerr"><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Nh1owl6a3HdaRB_Y0y04pw.png" /></a></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=55e7a5616198" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Voice Is the New Battleground]]></title>
            <link>https://medium.com/@batuhansenerr/voice-is-the-new-battleground-d06778e68e80?source=rss-2256a8221220------2</link>
            <guid isPermaLink="false">https://medium.com/p/d06778e68e80</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[news]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Batuhan Sener]]></dc:creator>
            <pubDate>Fri, 16 Oct 2026 15:57:05 GMT</pubDate>
            <atom:updated>2026-10-16T16:02:33.869Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KSrzfcFSn47kCvFHzFDpag.png" /></figure><p><em>Anthropic replaced git at 30,000 merges a day, Meta opened an 816B model, and Grok Voice 3.0 tops an independent speech index.</em></p><p>This issue covers September 28 to October 18, and every item in it was posted between October 11 and 15. Meta priced Muse Wonder, its first frontier model, the same as Claude Opus 5.5, and on the one independent index in Meta’s own table both Wonder models land a few points below it. Anthropic wrote up why it moved its monorepo off git once agents pushed merges past 30,000 a day, and its Frontier Red Team tested whether small open models can copy themselves onto cheap devices.</p><p><strong>In this issue:</strong> 2 top stories · 3 models · 4 tools · 6 quick hits · 1 quote</p><h3>Top Stories</h3><p><strong>Muse Wonder: Meta’s first frontier model costs the same as Opus 5.5 and scores a little lower</strong></p><p>Meta released Muse Wonder and Muse Wonder Full on October 12. Wonder is on the Meta Model API and Muse Code at $4 per million input tokens and $20 per million output tokens, which TechCrunch notes is the Opus 5.5 price. Wonder Full is only in Muse Code. Meta’s table leads with its own Personal Superintelligence Index, where Wonder Full scores 59.1 against 55.2 for Opus 5.5. The Artificial Analysis Index row at the bottom of the same table puts Opus 5.5 at 58, Wonder Full at 56 and Wonder at 52, with Fable 5.1 and GPT-6 Astra at 53. If you pay for Opus 5.5 today, the model you can call from Meta’s API gives you no price reason to switch, and the stronger one is not on the API.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Dx7E2ikOuNS__ZYkWRKitA.png" /><figcaption><em>The top of Meta’s benchmark table: Muse Wonder Full and Wonder against Fable 5.1, GPT-6 Astra and Opus 5.5.</em></figcaption></figure><p><a href="https://x.com/finkd/status/2109689305073614918">Source →</a></p><p><strong>Anthropic replaced git after agents pushed merges past 30,000 a day</strong></p><p>Anthropic merged about 2,500 code changes a day in May and more than 30,000 a day in September, and its git mirrors became the bottleneck. Its monorepo has close to 100 GB of history and serves about 25 million CI checkouts a day. A team of seven built Scion, a version control system where commits and merges happen on the server and clients fetch only the files they read, and moved the monorepo onto it in five months. Anthropic reports checkouts in seconds instead of minutes, an infrastructure footprint about 200 times smaller than the git mirrors, and 25% more CI CPU capacity. When agents write most of the code, merge queues, CI and review sized for humans become the slow part.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*mj6E7JjWUB-vgnNY.png" /></figure><p><a href="https://www.anthropic.com/engineering/redesigning-source-control">Source →</a></p><h3>New Models</h3><p><strong>Grok Voice 3.0: first on an independent speech-to-speech index, at the old price</strong></p><p>xAI released Grok Voice 3.0 in its Voice Agent API on October 14 at $0.08 per minute, the same as Grok Voice 2.0. Artificial Analysis scores it 86.9 on its Speech-to-Speech Index, ahead of GPT-Live-1 at 83.2 and Gemini 3.8 Live at 82.6. If you run a voice agent on Grok Voice 2.0, the upgrade costs nothing per minute, and xAI says agents on grok-voice-latest were switched automatically.</p><p><em>Heads up: The index scores are Artificial Analysis’. On τ-Voice, which tests support calls with accents, noise and interruptions, xAI’s own page has GPT-Live-1 ahead, 74.5% to 71.6%. The Starlink call-center figures on the page, about 70% of roughly 20,000 daily calls resolved without a human, are xAI’s own report.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*-9w8coT-gcabca09" /></figure><p><a href="https://x.com/SpaceXSI/status/2110416580899168305">Source →</a></p><p><strong>Muse Spark Open: 816B open weights that win on agent tasks and trail on coding</strong></p><p>Meta released Muse Spark Open on October 15, a mixture-of-experts model with 816B total and 42B active parameters, a 1M-token context and text, image, video and audio input. In Meta’s table it leads the open models listed on MCP-Atlas at 89.2 and JobBench at 61.8. On Terminal-Bench 2.1 it scores 76.4 against 85.0 for Kimi K3, and on DeepSWE 1.1 it scores 60 against 68. Pick it for tool-calling agents, but the table gives you no reason to use it for coding.</p><p><em>Heads up: Meta ran the comparison. The Muse Model License is not an open source license. Products must show “Built with Muse Spark Open”, and companies selling access to the model with more than $20M in revenue over 12 months need a separate agreement with Meta. The model card asks for at least 700GB of VRAM in NVFP4 and about 2TB in BF16.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qnAXpgtgYvIfQz98h68-iw.png" /><figcaption><em>Meta’s table: Muse Spark Open against Inkling, Nemotron 3 Ultra, Beam, Mistral Large 4, Qwen3.8, GLM 5.3 and Kimi K3.</em></figcaption></figure><p><a href="https://x.com/AIatMeta/status/2110779754274292212">Source →</a></p><p><strong>GLM-Legal: a legal model built on GLM-5.3-Flash, with gated access</strong></p><p>Z.ai released GLM-Legal on October 12 for contract review, evidence analysis and compliance work. On Harvey’s Legal Agent Benchmark its all-pass rate is 19.2%, against 12.9% for Kimi K3, 11.3% for Claude Fable 5, 6.7% for its GLM-5.3-Flash base and 5.4% for GPT-6 Astra. Domain training nearly tripled the base model’s score, though the best model on the chart still fails to fully pass more than 80% of the tasks.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*tJai2iV-pWJpmQAt" /></figure><p><a href="https://z.ai/blog/glm-legal">Source →</a></p><h3>Tools &amp; Libraries</h3><p><strong>UltraBERT</strong>: DMIS Lab’s 150M and 380M English encoders with a 32K context, plus an embedding model, under Apache 2.0, for classification, NER, QA, reward modeling and retrieval. Per @m_newhaus’s summary of the paper, the 380M leads their GLUE and BEIR comparison and the 150M matches RoBERTa-large on classification at 2.3x smaller. The training recipe uses 5% masking for classification and 50% for retrieval, with a separate LM head for each, and distills from a pruned 1.3B teacher only on the 5% sequences.</p><p>pip install -U transformers accelerate, then load dmis-lab/UltraBERT-380M with trust_remote_code=True</p><p><em>Who it’s for: anyone training or fine-tuning an encoder, even in another language, since the recipe carries over. The comparisons are the authors’ own.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*nhJ8HdHoL0jdmkdK" /></figure><p><a href="https://x.com/m_newhaus/status/2110393638618288257">Link →</a></p><p><strong>Fish Audio Drama 3</strong>: a text-to-speech model for multi-character scenes and long-form audio, with voice cloning and style control in plain language. It has 35B parameters with 3B active per token, and one text can carry up to 16 speakers. On an H200, median time to first audio is 88.9 ms with one stream and 450.9 ms with 64.</p><p>&lt;|speaker:0|&gt; ... &lt;|speaker:15|&gt; tags mark who says each line</p><p><em>Who it’s for: teams making audiobooks, dubbing or game dialogue. The card reports speed only, with no quality score such as word error rate, and it does not list its 83 languages.</em></p><p><a href="https://huggingface.co/fishaudio/drama-3">Link →</a></p><p><strong>Indra</strong>: a 1M-parameter OCR model from Vik Paruchuri that returns text with full layout and line bounding boxes, at 49 pages per second on a CPU, by his count. He reports 41.3 on OmniParseBench text, against 35.8 for PaddleOCR and 19.3 for Tesseract.</p><p><em>Who it’s for: pipelines that need fast OCR without a GPU. The numbers are his own, and the comparison covers only PaddleOCR and Tesseract.</em></p><p><a href="https://huggingface.co/datalab-to/indra-ocr">Link →</a></p><p><strong>Decision 3.0</strong>: six open decision models from vLLM Semantic Router, from 0.6B to 26B, listed on Hugging Face as zero-shot classifiers. Xunzhuo Liu says they rank first on both Jev Decision Index boards in the team’s own evaluation.</p><p><em>Who it’s for: routing and classification steps in agent pipelines, where the step only has to pick from a fixed list of options.</em></p><p><a href="https://huggingface.co/collections/vllm-sr/decision-30">Link →</a></p><h3>Quick Hits</h3><ul><li><strong>Gemini 4 Argon rolls out:</strong> Google says its frontier model, built for coding, finance, legal, cybersecurity and science, is rolling out globally in its products. <a href="https://x.com/GoogleAI/status/2110039860077527293">source</a></li><li><strong>Small models and self-replication:</strong> Anthropic’s Frontier Red Team found that a 9B model got inference running on 8 of 9 hardware types, and that GLM-5.2 spread across all 8 nodes of a test network in 4 of 5 trials. <a href="https://www.anthropic.com/research/self-replicating-ai">source</a></li><li><strong>QwenGyre:</strong> Qwen’s reinforcement learning framework for agent tasks with about 700K tokens per rollout lifted NL2RepoBench from 52.5% to 58.5% in 48 steps and cut end-to-end training time by about 44% against an asynchronous baseline, Qwen says. <a href="https://x.com/Alibaba_Qwen/status/2109633085658808515">source</a></li><li><strong>Maia 300:</strong> Satya Nadella says Microsoft’s next chip delivers 40% more tokens per watt than leading GPU systems, and that Maia will be the main inference system for Microsoft’s own products from next year. <a href="https://x.com/satyanadella/status/2109631106404487523">source</a></li><li><strong>GPT-Live-1.1:</strong> OpenAI’s voice update brings new voices, better accents and less “checking…” to ChatGPT, Codex and dots for all users today, with the API to follow. <a href="https://x.com/OpenAI/status/2110797830361436604">source</a></li><li><strong>FLUX 3 Image Klein:</strong> Black Forest Labs released a 5.6B open-weights image model that runs on a single consumer GPU, with native 4K, bounding boxes and pixel-perfect editing; the post does not name a license. <a href="https://x.com/bfl_ai/status/2110751201105256925">source</a></li></ul><h3>Quote of the Week</h3><blockquote>“Very often, we found that the role of the engineers was to tell Claude “no.””<em><br> — </em>Anthropic engineering blog, on how a team of seven built Scion with Claude</blockquote><h3>What To Do This Week</h3><p><strong>Try:</strong> Install Qwen-Image-2.1 from Hugging Face and run its image-editing example on one of your own photos. The steps follow the <a href="https://huggingface.co/Qwen/Qwen-Image-2.1">model card</a>.</p><p>Step 1: Install the libraries.</p><pre>pip install &quot;torch&gt;=2.4.0&quot;<br>pip install &quot;transformers&gt;=5.17&quot;<br>pip install git+https://github.com/huggingface/diffusers<br>pip install accelerate pillow</pre><p>Step 2: Save a photo as input.png and run the edit code<strong>.</strong> The first run downloads the model from Hugging Face.</p><pre>import torch<br>from PIL import Image<br>from diffusers import QwenImage21Pipeline<br><br>pipe = QwenImage21Pipeline.from_pretrained(<br>    &quot;Qwen/Qwen-Image-2.1&quot;, torch_dtype=torch.bfloat16<br>).to(&quot;cuda&quot;)<br><br>input_image = Image.open(&quot;input.png&quot;)<br><br>image = pipe(<br>    prompt=&quot;Change the background to a sunset beach&quot;,<br>    image=input_image,<br>    num_inference_steps=40,<br>    generator=torch.Generator(&quot;cuda&quot;).manual_seed(42),<br>).images[0]<br><br>image.save(&quot;edit_example.png&quot;)</pre><p>Step 3: Change the prompt and run it again. Qwen says the model keeps people and products recognizable and renders text better, so ask for an edit that adds a line of text and check both claims in edit_example.png.</p><p><strong>Read:</strong> Anthropic’s full write-up, which separates what it measured from what it extrapolated: <a href="https://www.anthropic.com/research/self-replicating-ai">AI models may be capable of self-replicating without access to a datacenter</a></p><p><strong>Watch:</strong> <a href="https://www.youtube.com/watch?v=_yHjTqistug">Let’s have some fun !!</a></p><p>If you would like to consult or get in touch, you can reach me via <a href="https://www.linkedin.com/in/yigitbatuhansener/"><strong>LinkedIn </strong></a>or <a href="http://ybatu.sener@gmail.com/"><strong>email</strong></a>.</p><p>If you enjoy my content, please consider showing your support by clapping for my articles or buying me a coffee. Your support is greatly appreciated.</p><figure><a href="https://buymeacoffee.com/batuhansenerr"><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Nh1owl6a3HdaRB_Y0y04pw.png" /></a></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d06778e68e80" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Decision Models, Cheaper Subagents and Europe’s Biggest Open Model]]></title>
            <link>https://medium.com/@batuhansenerr/decision-models-cheaper-subagents-and-europes-biggest-open-model-de5407b3c31b?source=rss-2256a8221220------2</link>
            <guid isPermaLink="false">https://medium.com/p/de5407b3c31b</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Batuhan Sener]]></dc:creator>
            <pubDate>Sun, 11 Oct 2026 13:45:45 GMT</pubDate>
            <atom:updated>2026-10-11T13:46:37.679Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Microsoft-Decision-1 and Unsloth’s training guide, Haiku 5.5 at $0.10 per million input tokens, and Mistral Large 4 at 1T parameters.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9S6VydHBRxvjoPV6IftVvw.png" /></figure><p>This issue covers September 21 to October 11, and every item that made it in was posted between October 6 and 9. Microsoft introduced Microsoft-Decision-1 and Unsloth published a guide to training your own decision model, which scores the options you give it and returns a choice with a probability. Anthropic priced Claude Haiku 5.5 at $0.10 per million input tokens and pitched it as a subagent, and Mistral put a 1 trillion-parameter model behind its API with open weights promised for the end of October.</p><p><strong>In this issue:</strong> 2 top stories · 3 models · 3 tools · 6 quick hits · 1 quote</p><h3>Top Stories</h3><p><strong>Mistral Large 4: 1T parameters, API now, open weights promised for October</strong></p><p>Mistral released Mistral Large 4 on October 6, a natively multimodal mixture-of-experts model with 1 trillion parameters. It is available through Mistral’s API today, and Mistral says the open weights follow at the end of October. Mistral calls it the best open-weights model from the US or Europe on aggregated benchmarks and says it trained it on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters. On DeepSWE 1.1, scored privately by Artificial Analysis, it reaches 62, ahead of GLM-5.3 at 61 and behind Kimi K3 at 68. On Dense200 visual grounding it scores 42.0 against 41.5 for GPT-6 Astra. If the weights ship on time, Europe gets a frontier-sized open model, though at 1T parameters few teams will host it themselves.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*TnjFZZwWPeXf1jre" /></figure><blockquote><strong><em>Heads up:</em></strong><em> Most numbers come from Mistral’s own announcement. The chart says Artificial Analysis ran DeepSWE 1.1 privately, ahead of the harness’s public launch, so the scores are not on its public index yet. The claim that it “surpasses closed frontier models on visual grounding” rests on a 0.5-point lead over GPT-6 Astra. Mistral’s own posts disagree on active parameters: 49B on X, 52 billion in the announcement. The weights are not out yet.</em></blockquote><p><a href="https://x.com/MistralAI/status/2107457414387622310">Source →</a></p><p><strong>Claude Haiku 5.5: Anthropic pitches its small model as the subagent</strong></p><p>Anthropic released Claude Haiku 5.5 on October 7 and describes it as “a fast, cost-efficient subagent for coding and well-defined tasks.” It costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100K tokens, and it is included in Claude Code. Anthropic’s chart puts it at 39.2% on Terminal-Bench 4.0, up from 0.0% for Haiku 4.5 but far behind Sonnet 5.5 at 70.6%, and at 72.4% on OSWorld 2.1 against 83.9% for Sonnet 5.5. That gap makes a split sensible. Let Haiku run the subagents that do lookups and narrow tasks, and keep the planning and general reasoning on Opus. Anthropic’s page says the same, with a model like Fable or Opus planning and handing subtasks to Haiku.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*uHCzTzv5Rwob8Zm9" /></figure><blockquote><a href="https://www.anthropic.com/claude/haiku">Source →</a></blockquote><h3>New Models</h3><p><strong>LightOnOCR-3: open OCR models from 0.8B that lead LightOn’s benchmark average</strong></p><p>LightOn released LightOnOCR-3 on October 8, a family of document models that read text and handwriting, describe images, pull data from charts and keep document structure in one pass. On LightOn’s average of olmOCR-Bench, ParseBench and fr-bench-pdf2md, the 4B model scores 78.5 and the 0.8B scores 76.9, against 75.0 for both Infinity-Parser2-Pro and chandra-ocr-2. With an Apache 2.0 license and a 0.8B version, you can run it on your own hardware inside a document pipeline.</p><p><em>Heads up: LightOn ran the benchmarks and publishes reproduction scripts on GitHub; nobody else has checked them yet. On olmOCR-Bench alone, Infinity Parser Pro scores 87.6 to the 4B’s 86.3, though Infinity Parser Pro has 35.1B parameters. The X post names Mistral OCR 4.1, while the chart shows Mistral OCR 4.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*6t5jbBACCh6PTlyD" /></figure><p><a href="https://x.com/LightOnIO/status/2108231119300645290">Source →</a></p><p><strong>Nano Banana 2.1: a Thinking variant tops Google’s own image evals</strong></p><p>Google DeepMind published the Nano Banana 2.1 model card on October 6. The model is built on Gemini 3.6 Flash and ships in the Gemini app, AI Studio, the Gemini API, Search AI Mode, Google Ads, Flow and Stitch. In Google’s side-by-side human evaluations, the Thinking variant scores an overall preference Elo of 1050 ± 14, against 990 ± 7 for Nano Banana 2, and its infographic factuality score, from an AutoRater, rises from 0.179 to 0.521. If you generate charts or diagrams, test the factuality jump on your own prompts before you trust it.</p><p><a href="https://deepmind.google/models/model-cards/nano-banana-2-1/">Source →</a></p><p><strong>Qwen-Image-2.1-Turbo: 2K images in 8 denoising steps</strong></p><p>Alibaba’s Qwen team released Qwen-Image-2.1-Turbo on October 9, an accelerated checkpoint on the same 7B architecture as Qwen-Image-2.1 that generates and edits images in 8 denoising steps. The weights are on Hugging Face and ModelScope, it loads in Diffusers through QwenImage21Pipeline, and Pro and Turbo APIs are live. Fewer steps means cheaper images on your own GPU, if the quality holds.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*aIKZbc_PJKuS4e_B.png" /></figure><p><a href="https://x.com/Alibaba_Qwen/status/2108549075218120949">Source →</a></p><h3>Tools &amp; Libraries</h3><p><strong>EmbeddingGemma 2</strong>: Google DeepMind’s open embedding model that maps text, code, images, video and audio into one 768-dimension space. It has 740M parameters in total, covers 100+ languages, reads up to 8,192 tokens and ships under Apache 2.0. On MTEB code it scores 78.68, up from 68.76 for the first EmbeddingGemma; on multilingual MTEB the gain is small, 61.36 against 61.15.</p><p><em>Who it’s for: teams building search or RAG over mixed media who want one small model instead of one per modality. The model card does not say who ran the benchmarks.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*E_uuiE_i98-94Dtn.png" /></figure><p><a href="https://huggingface.co/google/embeddinggemma-2">Link →</a></p><p><strong>Unsloth decision models</strong>: an open-source guide and notebooks for turning an LLM such as Qwen3.8 or Gemma 4 into a decision model that scores the options you give it and returns a choice with a probability instead of text. Unsloth fine-tuned with a Clef head and LoRA (r=64) for one epoch and raised Qwen3.5 0.8B’s aggregate accuracy from 20.7% to 74.3% across 3 decision benchmarks, on 4GB of VRAM.</p><p><em>Who it’s for: anyone curious about decision models, which are everywhere right now. The notebooks that show how to train one are the valuable part. Unsloth ran the numbers and says the test sets were decontaminated against the training data.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*gL2J3IgObcaZhQBr" /></figure><p><a href="https://x.com/UnslothAI/status/2107868866361930236">Link →</a></p><p><strong>MAI-Code-1.1-Flash in GitHub Copilot</strong>: Microsoft AI is bringing its MAI-Code-1.1-Flash model on-device, with local model calls inside GitHub Copilot and no inference charge for them.</p><p><em>Who it’s for: Copilot users who want agentic coding without paying per call. Microsoft’s post gives no benchmark or hardware requirements, so local quality is still unknown.</em></p><p><a href="https://x.com/MicrosoftAI/status/2107926901150908483">Link →</a></p><h3>Quick Hits</h3><ul><li><strong>HeyGen Voice tops an independent TTS arena:</strong> On Artificial Analysis’ Controlled Voice TTS Arena, HeyGen Voice holds an Elo of 1,201 across 1,468 appearances, ahead of Qwen-Audio-3.1-TTS-Plus at 1,182 and Eleven v4 Turbo at 1,166, at $30 per 1M characters. <a href="https://x.com/ArtificialAnlys/status/2108608255387935199">source</a></li><li><strong>GPT-6 reaches every ChatGPT user:</strong> OpenAI says GPT-6 and Intelligent UI, which answers with visuals and interactive tools instead of plain text, are rolling out in ChatGPT for everyone. <a href="https://x.com/OpenAI/status/2107894997538525580">source</a></li><li><strong>Gemini agent for business:</strong> Google Cloud announced a Gemini agent that runs 24/7 across your tools, data and work history, with built-in governance for data security and cost. <a href="https://x.com/googlecloud/status/2108252402075734172">source</a></li><li><strong>Microsoft-Decision-1:</strong> Satya Nadella introduced a decision model that Microsoft says beats LLMs and other decision models on latency and quality in structured decision tasks; the post gives no numbers. <a href="https://x.com/satyanadella/status/2108627923888754862">source</a></li><li><strong>TPUs in orbit:</strong> Google says it launched its first Project Suncatcher test satellite, carrying four TPUs, to find out whether its AI hardware can survive spaceflight, radiation and the temperature extremes of space. <a href="https://x.com/Google/status/2107875278416753052">source</a></li><li><strong>Google Labs Playground:</strong> An experimental platform for making your own games with no coding, open to users 18 and older in the US. <a href="https://x.com/GoogleLabs/status/2107800195748737042">source</a></li></ul><h3>Quote of the Week</h3><blockquote><em>“We’re already testing it across Microsoft for everything from incident response and quality control to scientific discovery.”<br> — Satya Nadella, Microsoft CEO, introducing Microsoft-Decision-1</em></blockquote><h3>What To Do This Week</h3><p><strong>Try:</strong> Split one task between a planner and a Haiku subagent in Claude Code, built on the Haiku 5.5 story above.</p><ol><li>Pick a task you repeat that has a clear lookup part, such as finding every caller of a function and summarising how each one uses it, and write down what “done” means.</li><li>Run it with your usual model doing everything. Note whether it met “done”, how long it took and the usage your plan reports.</li><li>Define a subagent in .claude/agents/ with model: haiku in its frontmatter and give it only the lookup part, while your main model plans and writes the summary.</li><li>Run the same task again and compare correctness, time and usage against the first run.</li><li>Keep the subagent if the result still meets “done” at lower usage; drop it if the main model has to redo the subagent’s work.</li></ol><ul><li><strong>Read:</strong> LightOn’s full write-up on LightOnOCR-3, with per-benchmark scores, speed tests and reproduction scripts: <a href="https://huggingface.co/blog/lightonai/lightonocr-3">LightOnOCR-3 on Hugging Face</a></li><li><strong>Watch:</strong> <a href="https://www.youtube.com/watch?v=BMLTQ0ouz1U">Let’s have some fun !</a></li></ul><p>If you would like to consult or get in touch, you can reach me via <a href="https://www.linkedin.com/in/yigitbatuhansener/"><strong>LinkedIn </strong></a>or <a href="http://ybatu.sener@gmail.com/"><strong>email</strong></a>.</p><p>If you enjoy my content, please consider showing your support by clapping for my articles or buying me a coffee. Your support is greatly appreciated.</p><figure><a href="https://buymeacoffee.com/batuhansenerr"><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Nh1owl6a3HdaRB_Y0y04pw.png" /></a></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=de5407b3c31b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Which Costs Less: a Better Model or a Better Harness?]]></title>
            <link>https://medium.com/@batuhansenerr/which-costs-less-a-better-model-or-a-better-harness-4d13ff926118?source=rss-2256a8221220------2</link>
            <guid isPermaLink="false">https://medium.com/p/4d13ff926118</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[news]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Batuhan Sener]]></dc:creator>
            <pubDate>Wed, 07 Oct 2026 10:40:29 GMT</pubDate>
            <atom:updated>2026-10-07T10:40:29.433Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*o9njRWpy3lEZ-lxm0eCLwA.png" /></figure><p><em>Artificial Analysis on Sonnet 5.5’s token bill, a harness that cuts agent costs by about half, and two weeks of open-weight releases.</em></p><p>Two weeks of releases, from September 21 to October 3. Artificial Analysis measured Anthropic’s Claude Sonnet 5.5 two points behind Opus 5.5 with the heaviest token use it has recorded, and an NVIDIA, NTU and MIT paper cut agent API costs by 50% to 54% against native harnesses. Xiaomi’s MiMo-V2.6-Pro took the top open-weight spot, and Google’s Gemini 4 Argon raised the output limit to 1M tokens.</p><p><strong>In this issue:</strong> 2 top stories · 3 models · 4 tools · 6 quick hits · 1 quote</p><h3>Top Stories</h3><p><strong>Sonnet 5.5 scores 56, two points behind Opus 5.5, and uses record tokens</strong></p><p>Artificial Analysis scored Anthropic’s Claude Sonnet 5.5 at 56 on its Intelligence Index, two points behind Opus 5.5 (max) at 58 and 18 points above Sonnet 5. The price matches Sonnet 5 at $2 and $10 per million input and output tokens, but the cost per task is $7.60, about 50% higher than Sonnet 5’s. At max effort it used about 193k output tokens per Index task, the most Artificial Analysis has measured: roughly 60% above Opus 5.5 (max) and about 7 times GPT-6 Astra (max). On Terminal-Bench 4.0 it reaches 64%, against 60% for Opus 5.5 and GPT-6 Astra. The same token price can hide a bigger bill, so compare models on cost per task.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vYlTue_aewJ1hvf5Ee6JBg.png" /><figcaption><em>Artificial Analysis Intelligence Index and cost per task</em></figcaption></figure><p><a href="https://x.com/ArtificialAnlys/status/2104640155843989864">Source →</a></p><p><strong>SoL-Pi cuts agent API costs by 50% to 54%, with a score drop on Opus 5</strong></p><p>Researchers at NVIDIA, NTU and MIT published SoL-Pi, an agent harness found by running Karpathy-style loops on the harness itself: propose a change, implement it, review it, test it, keep or discard. Four mechanisms survived. The harness fuses actions into one tool call, compacts context at completed steps, archives oversized outputs, and hands log reading to a cheaper model whose evidence is checked against the original. On the 51-task EdgeBench, API cost fell 50.0% against Codex with GPT-5.6 Sol ($1,787 to $894) and 54.3% against Claude Code with Opus 5 ($2,535 to $1,158). The harness set the bill.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*G3GRU0StlqzLlv7v5DpGOw.png" /><figcaption><em>Figure 1 from the SoL-Pi paper: the auto-research loop and held-out EdgeBench results.</em></figcaption></figure><p><a href="https://arxiv.org/pdf/2609.20519">Source →</a></p><h3>New Models</h3><p><strong>MiMo-V2.6-Pro: the top open-weight model on the Artificial Analysis Index</strong><br> Xiaomi released MiMo-V2.6-Pro, an open-weight mixture-of-experts model with 1.02T total and 42B active parameters. Artificial Analysis scores it 46 on its Intelligence Index, up from 26 for MiMo-V2.5-Pro and the highest among open-weight models, at $0.13 per Index task. API prices are $0.435 per million input tokens, with a 99% cache-hit discount, and $0.87 per million output tokens. A 1.02T-parameter model is not small to self-host, so the saving is mostly in the API price.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lwRiy7t-cJgJ6v_vYdEHEw.png" /><figcaption><em>Artificial Analysis Intelligence Index and cost per task,</em></figcaption></figure><p><a href="https://x.com/ArtificialAnlys/status/2102128560962187701">Source →</a></p><p><strong>Gemini 4 Argon: a 1M-token output limit, with access still narrow</strong><br> Google announced Gemini 4 Argon on September 30, aimed at long-horizon software engineering, enterprise knowledge work and cyber defense. The output limit rises from 64K to 1M tokens. Google reports 77.9% on DeepSWE v1.1, 51.3% on Zapier’s AutomationBench (ranked #1), 91.7% on LVBench, and a first-place tie at 68% on CWE-bench v1. It starts with cyber defenders and trusted testers, then paid API customers and Google AI Ultra subscribers, at an introductory $2 and $10 per million input and output tokens, rising to $4 and $20. Long outputs make long bills, so read it next to the Sonnet 5.5 token numbers above.</p><p><a href="https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/">Source →</a></p><p><strong>Kolibri: a 78B open mixture-of-experts model from Aleph Alpha</strong><br> Aleph Alpha released Kolibri, an open-weight model with 78B total and 3.46B active parameters, a 1M-token context window and an Apache 2.0 license. It has an explicit reasoning mode, tool calling and a bilingual German and English tokenizer, with 21.3% of pre-training tokens in German. Aleph Alpha says it was built with the EU AI Act, the General-Purpose AI Code of Practice and the GDPR in mind.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YL2MQFstVs_9Fv4PNLRO1g.png" /><figcaption><em>Aleph Alpha’s radar chart: Kolibri 78.1B-A3.46B against Kolibri Origin and three open models.</em></figcaption></figure><p><a href="https://huggingface.co/Aleph-Alpha/Kolibri-1">Source →</a></p><h3>Tools &amp; Libraries</h3><p><strong>Qwen Intelligence</strong>: Qwen’s personal intelligence product for phones, launched with three agents. A Mobile Planner Agent plans and orchestrates tasks, a Mobile-Use Agent works API-first with a GUI fallback, and a Mobile Creative Agent turns one sentence into an image in 3 seconds, which Qwen says is about 2x faster than leading peers. Qwen reports 82.1 on MobileWorld, 92.2 on MobileWorld-Real, 97.2 on AndroidDaily and 90% end-to-end success, and opened its benchmark suite.<br> qwenintelligence.com</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7PP-HtGGZbMx9mS6pdqb5g.png" /></figure><p><a href="https://www.qwenintelligence.com/">Link →</a></p><p><strong>GLiNER2.5-Decide</strong>: Fastino’s 340M-parameter, encoder-based open-weight model for fast classification and routing decisions. On Fastino’s own Fast Decisions suite of 17 datasets it scores 60.1%, ahead of JevK5 (57.5%), SemIf (56.4%) and Laya (46.6%), and it leads on 9 of the 17. Fastino says it runs on consumer CPUs and in air-gapped setups, under Apache 2.0.<br> huggingface.co/fastino/GLiNER2.5-Decide<br> <em>Who it&#39;s for: teams routing requests or classifying text without sending data out. The suite is Fastino&#39;s own, internally generated, and the lead over the next model is 2.6 points.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zyC0pzITESJgDgUbCRReRg.png" /></figure><p><a href="https://huggingface.co/fastino/GLiNER2.5-Decide">Link</a> →</p><p><strong>lift</strong>: Datalab’s open-weight 9B model that takes a PDF or image plus a JSON schema and returns a matching JSON object in one pass, reading every page so values that span pages stay right. The thread reports 90.2% field accuracy on 11,000 fields, self-hosting with vLLM or Hugging Face, and a null instead of a guess when a value is missing.<br> github.com/datalab-to/lift<br> <em>Who it&#39;s for: teams extracting fields from invoices or papers. </em><br><a href="https://github.com/datalab-to/lift"> Link →</a></p><p><strong>Decision-2.0</strong>: a 7.94B decision model from the vLLM Semantic Router team, built on Qwen3.5 and released under Apache 2.0. It answers choice, yes/no and scoring questions in one pass and returns probabilities without generating text. It scores 68.1 on JevArena against 65.8 for Decision-1.0-Lux and 62.1 for Nimble v2.<br> <em>Who it&#39;s for: teams routing or classifying requests. The scores are on the card&#39;s own suites, and the gap on human-labelled transfer is small, 56.2 against 55.8.</em> <br> <a href="https://huggingface.co/collections/vllm-sr/decision-20">Link →</a></p><h3>Quick Hits</h3><ul><li><strong>OpenAI dots:</strong> OpenAI introduced dots, always-on agents powered by GPT-6 Astra, in a short post with a video. Secondary coverage says each dot works from its own cloud computer and browser and is rolling out to Pro and Business Premium subscribers, but OpenAI’s own page could not be read, so treat those details as secondhand.<a href="https://help.openai.com/en/articles/20001530-getting-started-with-your-dot"><br>source</a></li><li><strong>Nemotron-3-Diarization:</strong> NVIDIA’s 100M-parameter open-weight speaker diarization model, under the OpenMDW License 1.1, handles up to 8 speakers in streaming or offline mode. At a 30.4-second buffer it scores 12.73% diarization error on DIHARD III against 19.09% for the card’s baseline, and 6.40% on AliMeeting Near against 11.57%. The listed languages are English, Mandarin, Hindi, Kannada, Telugu and Bengali. <br><a href="https://huggingface.co/nvidia/Nemotron-3-Diarization">source</a></li><li><strong>Claude and a new enzyme system:</strong> Anthropic says Claude found a previously unknown enzyme system in bacteriophage DNA, with a long array of repeating DNA beside its gene that looks somewhat like CRISPR. What it does is not yet known, though the few known systems with these features can all cut, copy and paste DNA.<br> <a href="https://x.com/AnthropicAI/status/2102824959827742916">source</a></li><li><strong>Grok 4.7:</strong> Elon Musk wrote that Grok 4.7 places SpaceXAI third, after Anthropic and OpenAI, for agentic coding, and that it is faster and cheaper. The post names no benchmark, and on Artificial Analysis’ general Intelligence Index Grok 4.7 (xhigh) sits at 46, below Opus 5.5 at 58.<br> <a href="https://x.com/elonmusk/status/2102082011233931762">source</a></li><li><strong>Microsoft AI’s transcription claim:</strong> Mustafa Suleyman said Microsoft AI is launching the most accurate real-time transcription model in the world: #1, 55% faster and 60% cheaper than ElevenLabs. The post names no model, ranking or language set, so there is nothing to check yet. <br><a href="https://x.com/mustafasuleyman/status/2105699115602677984">source</a></li><li><strong>LeCun on scaling:</strong> A summary of Yann LeCun’s ETH Zürich talk, posted by Rohan Paul, has him calling AGI by scaling LLMs “impossible”. The argument: about 30 trillion training tokens, roughly 10¹⁴ bytes, would take a person around 400,000 years to read, while a 4-year-old takes in the same amount through vision in about 1 year and 10 months. It is a secondhand summary of one researcher’s view.<br> <a href="https://x.com/rohanpaul_ai/status/2106337846198214728">source</a></li></ul><h3>Quote of the Week</h3><blockquote>“As coding agents move from supervised code completion to unattended, around-the-clock exploration, their work expands from isolated predictions into long trajectories of reasoning, tool use, and feedback. Token efficiency therefore becomes important for scaling recursive self-improvement.”<br>— Haozhe Liu and coauthors, NVIDIA, NTU and MIT, SoL-Pi paper (arXiv 2609.20519)</blockquote><h3>What To Do This Week</h3><ul><li><strong>Try</strong>: A small cost-per-task experiment in Claude Code, built on the two top stories above.<br> 1. Pick one task you repeat and can score pass or fail, and write down what “done” means.<br> 2. Run it three times with each of two models, noting whether it finished, retries, tokens or cost, and time.<br>3. Divide each model’s total cost by its finished runs. A cheaper model that fails more often can cost more per finished task.<br>4. Add one change from SoL-Pi: let a subagent on a cheaper model do the searching and log reading, and give the main model only a summary you can check against the original.<br>5. Run it again. Keep the change only if the cost per finished task drops and success holds.</li><li><strong>Read:</strong> SoL-Pi’s four harness mechanisms, if you run coding agents: <a href="https://arxiv.org/abs/2609.20519">arXiv 2609.20519</a></li><li><strong>Watch:</strong> <a href="https://www.youtube.com/watch?v=jJBD1EzCam8&amp;list=RDdKFCr_mrXS4&amp;index=2">Let’s have some fun !</a></li></ul><p>If you would like to consult or get in touch, you can reach me via <a href="https://www.linkedin.com/in/yigitbatuhansener/"><strong>LinkedIn </strong></a>or <a href="http://ybatu.sener@gmail.com/"><strong>email</strong></a>.</p><p>If you enjoy my content, please consider showing your support by clapping for my articles or buying me a coffee. Your support is greatly appreciated.</p><figure><a href="https://buymeacoffee.com/batuhansenerr"><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Nh1owl6a3HdaRB_Y0y04pw.png" /></a></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4d13ff926118" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Fable 5.1 Hits 90% on ARC-AGI-2, and Amodei Asks to Slow Down]]></title>
            <link>https://medium.com/@batuhansenerr/fable-5-1-hits-90-on-arc-agi-2-and-amodei-asks-to-slow-down-03b5c865f3be?source=rss-2256a8221220------2</link>
            <guid isPermaLink="false">https://medium.com/p/03b5c865f3be</guid>
            <category><![CDATA[news]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Batuhan Sener]]></dc:creator>
            <pubDate>Sat, 03 Oct 2026 17:34:11 GMT</pubDate>
            <atom:updated>2026-10-03T17:34:11.882Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6ExTUYz5XUoxpuviQb263g.png" /></figure><p>Three weeks of releases, from August 31 to September 20. ARC Prize verified Anthropic’s Fable 5.1 at 90.0% on ARC-AGI-2, and Dario Amodei published an essay arguing that labs should slow the pace at which they improve capabilities. Meta, DeepSeek and Qwen shipped models with their own benchmark numbers.</p><p><strong>In this issue:</strong> 2 top stories · 4 models · 1 tools · 5 quick hits · 1 quote</p><h3>Top Stories</h3><p><strong>Fable 5.1 reaches 90.0% on ARC-AGI-2, for less money than Fable 5</strong></p><p>ARC Prize verified Anthropic’s Fable 5.1 at 90.0% on ARC-AGI-2 for $3.12 per task, and at 97.5% on ARC-AGI-1 for $1.40 per task. Averaged over the two benchmarks, ARC Prize says cost per task is about 32% below Fable 5’s, which it credits to better token efficiency. On ARC Prize’s chart, Fable 5 sits just below Fable 5.1, so the saving did not cost accuracy. About a third off per task at the same score is a gain you can put in a budget.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*woyc6pqfk7G8YAl8pIgA8A.png" /></figure><p><em>ARC Prize’s ARC-AGI-2 leaderboard: score against cost per task, with Claude Fable 5.1 and Fable 5 highlighted.</em></p><blockquote><strong><em>Heads up:</em></strong><em> ARC Prize ran the test, not Anthropic, which makes this stronger than a vendor claim. The chart plots score against cost per task, and several other models plot close to the top of it, a GPT-5.6 Sol entry among them.</em></blockquote><p><a href="https://arcprize.org/results/anthropic-claude-fable-5-1">Source</a> →</p><p><strong>Amodei asks AI labs to slow down, and puts Anthropic’s access rules on the table</strong></p><p>On September 12, Dario Amodei announced an essay titled “We Must Pace the Frontier.” Its central sentence: “We must slow the pace at which we improve the capabilities of AI models.” Pacing, he says, does not mean stopping training. His plan has three parts: independent evaluators embedded inside AI companies, shared safety standards among democracies, and an attempt at coordination with authoritarian states.</p><p>Anthropic is committing to the first part on its own. Third-party evaluators would get permanent, employee-level access to its systems and the right to publish findings without editorial control by Anthropic, which keeps narrow redaction rights for security-sensitive, legally privileged, commercially sensitive or third-party confidential material. Desks and badges are easy to promise. Findings published without Anthropic’s editing are what outsiders can check.</p><p><a href="https://darioamodei.com/post/we-must-pace-the-frontier">Source →</a></p><h3>New Models</h3><p><strong>Muse Spark 1.3: a long-context lead, and fewer tool calls</strong><br> Meta released Muse Spark 1.3, aimed at agentic and coding work. In what Meta calls internal comparisons, it uses about 20% fewer tool calls and about 25% fewer tokens than Muse Spark 1.2. Meta’s table puts it at 98.1 on MRCR 512K-1M, against 73.8 for GPT 5.6 Sol and 55.5 for Muse Spark 1.2; Opus 5 has no score listed. Opus 5 leads four of the six agent benchmarks, including OSWorld 2.0 (68.3 vs 66.9), and GPT 5.6 Sol leads DeepSearchQA (93.0 vs 89.4). If you push long documents through a model, test the long-context number on your own data.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*k8UpaNnKtLN_cK1Z1NwG2A.png" /><figcaption><em>Meta’s benchmark table: Muse Spark 1.3 against Muse Spark 1.2, GPT 5.6 Sol and Opus 5.</em></figcaption></figure><p><a href="https://research.meta.ai/blog/introducing-muse-spark-1-3">Source →</a></p><p><strong>DeepSeek-V4.1-Flash: level with Opus 5 on coding, behind on terminal work</strong><br> DeepSeek released V4.1-Flash, the smallest model in its new architecture family, with native visual understanding. DeepSeek’s chart has it at 74.2 on DeepSWE v1.1 (Opus 5: 74.0), 88.1 on CyberGym (84.5 for GLM-5.3 and GPT5.6-Sol) and 54.8 on Automation-Bench (Opus 5: 50.3). On Terminal-Bench 3.0 it trails, with 30.0 against 43.3 for Opus 5 and 34.4 for GPT5.6-Sol. A Flash-size model within 0.2 points of Opus 5 on DeepSWE is a candidate for your own coding evals.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9YaiD5zzylHttlQniEAQbQ.png" /><figcaption><em>DeepSeek’s chart: V4.1-Flash against Kimi-K3, GLM-5.3, Opus 5 and GPT5.6-Sol on four benchmarks.</em></figcaption></figure><p><a href="https://x.com/deepseek_ai/status/2097930608790167907">Source →</a></p><p><strong>Qwen-Image-2.1</strong>: Qwen’s image generation and editing model, now released with open weights. It is a 7B model that takes up to 10 reference images and generates and edits RGBA layers natively, which helps if you composite or edit text inside transparent images.<br> <em>Who it’s for: teams that want a self-hosted image editor. Qwen says it outperforms most closed-source models and the post names no benchmark.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*z9Az2YdAzvqAMzMDb3jxGw.png" /><figcaption><em>Qwen’s announcement graphic for Qwen-Image-2.1: “Now open weights.”</em></figcaption></figure><p><a href="https://huggingface.co/Qwen/Qwen-Image-2.1">Link</a> →</p><p><strong>Qwen3.8-LiveTranslate</strong>: Qwen’s real-time simultaneous interpretation model. Qwen says it cuts average lag (LAAL) from 2.8 to 2.3 seconds across 60 languages, tells speakers apart in multi-party speech while keeping each voice, and shows source and translation side by side.<br> qwen3.8-livetranslate-flash-realtime<br> <em>Who it&#39;s for: teams building live captioning or interpretation. The post gives one average across 60 languages, so test the ones you need, and it links an API page but no weights.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uXmBr2RJ5wGe8r1zh54YEw.png" /></figure><p><a href="https://www.qwencloud.com/models/qwen3.8-livetranslate-flash-realtime">Link →</a></p><h3>Tools &amp; Libraries</h3><p><strong>zg</strong>: a local search tool for developers and AI agents, open-sourced by Alibaba’s Zvec team. It is local-first, works out of the box with popular agents, and combines semantic, BM25, hybrid and rg search in one tool.<br> <em>Who it’s for: anyone connecting a coding agent to a local codebase. Why the team built it is in an X Article that could not be read, so this is Qwen Devs’ summary only.</em> <br> <a href="https://github.com/zvec-ai/zvec-grep">Link</a> →</p><h3>Quick Hits</h3><ul><li><strong>WeatherNext 3:</strong> Google says forecasts are up to 5x sharper than WeatherNext 2. It takes live geostationary satellite observations as inputs and updates the global forecast every hour, where traditional physics-based models can carry a 6-hour lag. Live observations replace a batch simulation cycle as the input. <a href="https://x.com/GoogleAI/status/2095544944190788064">source</a></li><li><strong>TimesFM-3:</strong> Google Research’s time series foundation model does multivariate forecasting in a single forward pass and, Google says, significantly outperforms other forecasting models across major benchmarks. The post names none of them, so test it on your own series. <a href="https://x.com/GoogleResearch/status/2094483372718580066">source</a></li><li><strong>Google Pics:</strong> Google’s AI image tool is rolling out to Workspace customers and Google AI Pro and Ultra subscribers at pics.new. It edits individual objects, refines or translates text, and lets teams collaborate, so the feature list reads like an editor’s. <a href="https://x.com/GoogleWorkspace/status/2094818432973725961">source</a></li><li><strong>Qwen-Drive-1.0:</strong> an Apache 2.0 model for driving question answering, 3D perception and motion planning, built on an unmodified Qwen3.5–4B. On Qwen’s chart it scores 77.8 on LingoQA (MiMo-Embodied-7B: 72) and 90.7 on NAVSIM PDMS (ExploreVLA: 90.4). The post says other models score under 5 on causal reasoning, where Qwen-Drive gets 41.3, but the same chart shows Cosmos-Reason2–32B at 5.73. <a href="https://x.com/ModelScope2022/status/2095357812620603729">source</a></li><li><strong>McKinsey HR Monitor:</strong> HR professionals naming digital literacy a top-five future skill rose 15 points from 2025 to 2026; problem solving stayed the most-cited at about 44% (up 9 points); digital learning fell 8 points. The survey measures perception, and 2026 added Belgium, China and the Netherlands, so McKinsey calls the year-on-year changes directional only.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*K6iyDSLbty_lVK5l6kkujA.png" /></figure><h3>Quote of the Week</h3><blockquote><em>“To be clear, pacing does not mean halting model training or technical progress, but ensuring companies take adequate time to align and safeguard their models, and for third party evaluators to confirm this.”<br> — Dario Amodei, CEO of Anthropic, “We Must Pace the Frontier”</em></blockquote><h3>What To Do This Week</h3><ul><li><strong>Try:</strong> Open Claude Code and write a skill for one task you repeat every week …</li><li><strong>Read:</strong> The evaluator access commitments in Amodei’s essay, “We Must Pace the Frontier”: <a href="https://darioamodei.com/post/we-must-pace-the-frontier">darioamodei.com</a></li><li><strong>Watch:</strong> Your cool-down after all those benchmark charts… let’s have some fun. <a href="https://www.youtube.com/watch?v=0UlMuCU-eI4&amp;list=RD0UlMuCU-eI4&amp;start_radio=1">No evals required.</a></li></ul><p>If you would like to consult or get in touch, you can reach me via <a href="https://www.linkedin.com/in/yigitbatuhansener/"><strong>LinkedIn </strong></a>or <a href="http://ybatu.sener@gmail.com/"><strong>email</strong></a>.</p><p>If you enjoy my content, please consider showing your support by clapping for my articles or buying me a coffee. Your support is greatly appreciated.</p><figure><a href="https://buymeacoffee.com/batuhansenerr"><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Nh1owl6a3HdaRB_Y0y04pw.png" /></a></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=03b5c865f3be" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Future of Investing — AI-Powered Financial Analysis with Finciples.ai]]></title>
            <link>https://medium.com/@batuhansenerr/the-future-of-investing-ai-powered-financial-analysis-with-finciples-ai-08a21140dd41?source=rss-2256a8221220------2</link>
            <guid isPermaLink="false">https://medium.com/p/08a21140dd41</guid>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[investing]]></category>
            <category><![CDATA[fintech]]></category>
            <category><![CDATA[finance]]></category>
            <dc:creator><![CDATA[Batuhan Sener]]></dc:creator>
            <pubDate>Thu, 13 Feb 2025 18:55:36 GMT</pubDate>
            <atom:updated>2025-02-13T18:55:36.718Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>The Future of Investing — AI-Powered Financial Analysis with Finciples.ai</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OxLz8jWPOTZ4lmqr2SR4pw.png" /></figure><p>In developing <a href="https://finciples.ai/"><strong>Finciples.ai</strong></a>, we aimed to provide investors with more than just market data. <br>By combining traditional technical and fundamental analysis with time-tested investment principles, including user’s custom strategies, the platform offers AI-powered insights and a comprehensive market view. Finciples.ai not only simplifies access to professional-grade analysis but also <strong>Augments Decision-Making</strong>, helping investors make more informed, principle-driven decisions.</p><h4>Who is Finciples.ai for ?</h4><p><strong>Individual Retail Investors</strong>: Investors seeking professional analysis and those aligning their portfolios with investment philosophies.</p><p><strong>Financial Professionals</strong>: Financial advisors strengthening client communication and portfolio managers aligning with specific philosophies.</p><p><strong>Educational Users</strong>: Students and educators using the platform to learn different investment approaches.</p><h4>Key Features</h4><p>The key features to be explored in this article.</p><ul><li><strong>Traditional Analysis</strong></li><li><strong>Principle Analysis — AI Augmented</strong></li><li><strong>Principle Analysis — Led</strong></li><li><strong>Analysis Transactions</strong></li><li><strong>Settings</strong></li></ul><h4><strong>Traditional Analysis</strong></h4><p>This module combines fundamental and technical analysis to enhance investment decisions. It displays trends with real-time stock charts, analyzes market movements using moving averages and volume indicators. It measures investor sentiment with social media sentiment analysis from Reddit and evaluates a company’s strength using key metrics like P/E ratio and market value. Using all these data, it provides AI-powered insights into stocks.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xJ_3H_KakYHP18t6UAHRaA.gif" /></figure><h4>Principle Analysis — <strong>AI Augmented</strong></h4><p>Principle Analysis is an AI-powered analytics platform that allows investors to create personalized investment principles and evaluate stocks accordingly. <br>Users can conduct stock analyses using different AI models based on their defined investment principles. <br>The platform utilizes a <strong>Mixture-of-Experts (MoE)</strong> architecture, integrating technical, fundamental, and sentiment analyses, along with investment legends’ principles and personalized principles. Through its scoring algorithms, each stock is rated according to predefined criteria, guiding investment decisions. <br>AI-powered analysis reports provide comprehensive insights, accelerating the decision-making process.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bUUdrok9StKW3j7_axdSvA.gif" /></figure><h4>Principle Analysis — AI Led</h4><p>AI-led Principle Analysis generates investment principles based on the selected language model and conducts stock analysis accordingly. <br>It evaluates stocks against these principles and provides scoring insights to guide investment decisions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sBdecw4vufEIWXgF55pozg.gif" /></figure><h4><strong>Analysis Transactions</strong></h4><p>In the Transaction section, users can save AI-supported evaluations based on both traditional and principle analysis for comparative archiving. Each transaction record includes the stock symbol, analysis date, investment philosophies used, AI model preferences, and detailed compatibility scores.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3C9R_9GqaITYQWzdjVIVXg.gif" /></figure><h4><strong>Settings</strong></h4><p>In <strong>Principles Settings</strong>, you can create your own investment principles and integrate those of renowned investors into your custom advisor model.<br>You also have the option to share your custom advisors with other users. The “Similarity” feature allows you to identify and cluster similarities between your investment strategies and those of other investors.</p><p><strong>Profile Settings</strong> personalizes stock analysis recommendations based on the data you provide.<br>For example, Warren Buffett’s investment strategy may generate different recommendations for AAPL stock depending on whether your goal is long-term or short-term. Likewise, investors with different risk tolerances may receive tailored insights. <br>This approach enhances decision-making by leveraging AI-driven personalization.</p><p>With <strong>Watchlist Settings</strong>, can create and manage personalized watchlists, allowing you to analyze selected stocks using Principle Analysis.</p><p><strong>LLM Settings</strong> provides access to integrations with the latest Large Language Models (LLMs). Depending on your user tier, you can choose multiple LLMs for a more comprehensive and generalized analysis across different stocks.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bWzOlee3_Ow-cT_XTrwykQ.gif" /></figure><p>Gain deeper insights into the relationships between users, stocks, and advisors through interactive dashboards that provide a comprehensive analytical view.<br>Engage with stock analysis in a more dynamic way through audio and video content listen to insights or watch key market trends instead of just reading. And this is just the beginning more innovative features are on the way.</p><h4><strong>AI Technologies in Finciples.ai</strong></h4><ul><li>The similarity calculation is based on an AI model that converts texts into vectors to interpret investment principles. <br>It utilizes the <strong>All-MiniLM-L6-v2</strong> model for vectorization, cosine similarity for comparison, weighted embedding for importance, and the <strong>K-Means</strong> algorithm for clustering similar investment advisors.</li><li>The chatbot is built using <strong>Langflow</strong> with an <strong>Agentic RAG system</strong>, enabling dynamic information retrieval, contextual understanding, and adaptive decision-making for more accurate and intelligent responses.</li><li>The fine-tuned on financial data based on <strong>DistilRoBERTa </strong>performs sentiment analysis on financial comments, classifying them as negative, neutral, or positive. Texts are processed using a tokenizer, sentiment scores are calculated using a softmax activation function, and insights supporting investment decisions are generated.</li><li>Chain of Thought techniques, Mixture of Experts, integrations of various language models, AI applications such as text-to-speech and text-to-video conversion, and much more…</li></ul><h4><a href="https://finciples.ai/"><strong>You can use the link to try the f</strong>inciples.ai</a>.</h4><p>If you would like to consult or get in touch, you can reach us via <a href="https://www.linkedin.com/in/yigitbatuhansener/"><strong>LinkedIn </strong></a>or <a href="http://finciples.ai@gmail.com/"><strong>email</strong></a>.</p><p>Thank you for reading this far. If you enjoy my content, please consider showing your support by clapping.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=08a21140dd41" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[YOLO11 Custom Object Detection]]></title>
            <link>https://medium.com/@batuhansenerr/yolo11-custom-object-detection-2116db587f3b?source=rss-2256a8221220------2</link>
            <guid isPermaLink="false">https://medium.com/p/2116db587f3b</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[object-detection]]></category>
            <category><![CDATA[yolov11]]></category>
            <category><![CDATA[computer-vision]]></category>
            <dc:creator><![CDATA[Batuhan Sener]]></dc:creator>
            <pubDate>Tue, 01 Oct 2024 14:39:31 GMT</pubDate>
            <atom:updated>2024-10-01T14:39:31.662Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Fewer parameters, faster, and more accuracy …</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/1*K-q7JKp232OmvA2mDG9bxw.jpeg" /></figure><h4>Overview</h4><p>YOLO11 was announced at the YOLO Vision 2024 event on September 27, 2024. If you missed the event, I highly recommend watching it, as there were some very valuable speakers in the fields of ML and CV. <a href="https://www.youtube.com/watch?v=rfI5vOo3-_A">Here’s the link for you to check it out.</a></p><p>YOLO11 is the latest version of the Ultralytics YOLO series, combining cutting-edge accuracy, speed, and efficiency for object detection, segmentation, classification, oriented bounding boxes, and pose estimation.</p><p>With fewer parameters and better results compared to YOLOv8, it’s not hard to foresee that YOLO11, being more efficient and faster on edge devices, will frequently appear as a SOTA in the field of computer vision.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*88ORyUAW1tqwhCVjCKY-ew.png" /></figure><h4>Key Features</h4><ul><li><strong>Enhanced Feature Extraction:</strong> YOLO11 uses an improved backbone and neck architecture to enhance feature extraction for more precise object detection and complex task performance.</li><li><strong>Optimized for Efficiency and Speed:</strong> Refined architectural designs and optimized training pipelines deliver faster processing speeds while maintaining an optimal balance between accuracy and performance.</li><li><strong>Greater Accuracy with Fewer Parameters:</strong> YOLO11m achieves higher mAP on the COCO dataset using 22% fewer parameters than YOLOv8m, increasing computational efficiency without compromising accuracy.</li><li><strong>Adaptability Across Environments:</strong> YOLO11 can be seamlessly deployed across edge devices, cloud platforms, and systems with NVIDIA GPUs, ensuring maximum flexibility.</li><li><strong>Broad Range of Supported Tasks:</strong> YOLO11 supports various computer vision tasks such as object detection, instance segmentation, image classification, pose estimation, and oriented object detection (OBB).</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fp9c61fuVMwdYGhVvhNSwg.png" /><figcaption><a href="https://leaderboard.roboflow.com/">Before starting the project, let’s look at the high accuracy with a low number of parameters on the Roboflow Leaderboard compared to other models.</a></figcaption></figure><h4>Steps Covered in this Tutorial</h4><ul><li>Environment Setup</li><li>Preparing the dataset</li><li>Training the model</li><li>Validate the model</li><li>Running inference on test images</li><li>Conclusion</li></ul><h4>Environment Setup</h4><p>You need a google account to use Google Colab.<br>We use Colab for tasks that require intensive computation, such as deep learning.<br>Since my computer’s GPU is insufficient, I need to activate Colab’s GPU support.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*icW292rtYjZGlPk3zxoplQ.gif" /></figure><p>After doing this, we check the gpu activity.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/690/1*EwqOikkgtj_Y13ObuBnpJg.png" /></figure><p>It supports up to 16GB of memory and 2560 CUDA cores to accelerate a wide range of modern applications.</p><p>And then execute this code to dynamically determine the working directory and manage file paths flexibly.</p><pre>import os<br>HOME = os.getcwd()</pre><p>Next, you need to download the Ultralytics package to load and work on models, and the Roboflow package for the dataset.</p><pre>!pip install ultralytics supervision roboflow<br><br>from ultralytics import YOLO<br>from roboflow import Roboflow</pre><h4>Preparing the Dataset</h4><p>In this project, I used the <a href="https://universe.roboflow.com/roboflow-100/parasites-1s07h"><strong>parasites</strong> </a>dataset included in <a href="https://universe.roboflow.com/roboflow-100"><strong>RF100</strong></a>.</p><p>I will train an object detection model with 8 different parasite classes in this dataset. I will work on labeled, classified images through Roboflow. I frequently use this open-source platform in my individual projects. When working with ready-made datasets, you can quickly access a lot of information about the data in the <a href="https://universe.roboflow.com/roboflow-100/parasites-1s07h/health"><strong>dataset’s health analytics</strong></a> section. For example, in the <strong>Class Balance</strong> section below, we can see that the Hymenolepis class is underrepresented.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*L4tgohtxT7NIGUYODP4lYw.png" /></figure><p>To improve accuracy for this class, you would need to apply augmentation, oversampling, or make adjustments to class weights. We won’t go into these topics in this article, don’t worry, but if you’re interested in these types of tasks, feel free to contact me. If there’s enough demand, I can also share my detailed work on these subjects.</p><p>You can download and use any of the open source projects in the roboflow environment in accordance with the format. After preparing or selecting the data set, we will work in the Colab environment.<br>The reason we switched to Colab is that it provides 16GB free NVIDIA T4 GPU usage. I touch upon this issue below.</p><p>We need to import our data in YOLOv8 format using the Roboflow API. Let’s get the ready code snippet and data format.</p><pre>rf = Roboflow(api_key=&quot;your-api-key&quot;)<br>project = rf.workspace(&quot;roboflow-100&quot;).project(&quot;parasites-1s07h&quot;)<br>version = project.version(2)<br>dataset = version.download(&quot;yolov8&quot;)<br>!sed -i &#39;$d&#39; {dataset.location}/data.yaml   # Delete the last line<br>!sed -i &#39;$d&#39; {dataset.location}/data.yaml   # Delete the second-to-last line<br>!sed -i &#39;$d&#39; {dataset.location}/data.yaml   # Delete the third-to-last line<br><br>!echo &#39;test: ../test/images&#39; &gt;&gt; {dataset.location}/data.yaml<br>!echo &#39;train: ../train/images&#39; &gt;&gt; {dataset.location}/data.yaml<br>!echo &#39;val: ../valid/images&#39; &gt;&gt; {dataset.location}/data.yaml</pre><p>We need to update the <strong>data.yaml</strong> format as shown below for training in YOLO11 format.</p><pre>!sed -i &#39;$d&#39; {dataset.location}/data.yaml   # Delete the last line<br>!sed -i &#39;$d&#39; {dataset.location}/data.yaml   # Delete the second-to-last line<br>!sed -i &#39;$d&#39; {dataset.location}/data.yaml   # Delete the third-to-last line<br><br>!echo &#39;test: ../test/images&#39; &gt;&gt; {dataset.location}/data.yaml<br>!echo &#39;train: ../train/images&#39; &gt;&gt; {dataset.location}/data.yaml<br>!echo &#39;val: ../valid/images&#39; &gt;&gt; {dataset.location}/data.yaml</pre><p><strong>Training the Model</strong></p><p>Let’s train a model on our dataset for 40 epochs.</p><p>As part of this project, I demonstrated the training through CLI commands. After making the definitions with a simple command as shown below, the training begins.</p><pre>!yolo task=detect mode=train model=yolo11s.pt data={dataset.location}/data.yaml epochs=40 imgsz=640 plots=True</pre><p>You can also create it for your own projects using Python, allowing for more flexible scenarios. Example Python training code is shown below :</p><pre>from ultralytics import YOLO<br><br># Load a COCO-pretrained YOLO11n model<br>model = YOLO(&quot;yolo11n.pt&quot;)<br><br># Train the model on the COCO8 example dataset for 40 epochs<br>results = model.train(data=&quot;coco8.yaml&quot;, epochs=40, imgsz=640)</pre><p>The trained model is saved as best.pt under /runs/detect/train/weights file.<br>Google Colab notebooks have an idle timeout of 90 minutes and absolute timeout of 12 hours.<br>If you don’t want your data to be lost, back it up !</p><p>Once your model has finished training, you can evaluate the training results using the graphs generated by the YOLO11.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/655/1*rfRDPfiAjIBvY5j4EA4pMg.png" /><figcaption>Performance Metrics</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/586/1*Ali-Z1D2GUb8BZd1jcHVAg.png" /><figcaption>Confusion Matrix</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/659/1*xO0fk0BDidwgWShJxRPH0w.png" /></figure><p><strong>Note:</strong> When training the models, it is beneficial to review the parameters in the <a href="https://docs.ultralytics.com/modes/train/#resuming-interrupted-trainings"><strong>Train Setting section of the Ultralytics documentation</strong></a>. This part could be crucial for your training processes.</p><h4>Validate the model</h4><p>Here’s why using YOLO11’s Val mode is advantageous:</p><ul><li><strong>Precision:</strong> Get accurate metrics like mAP50, mAP75, and mAP50–95 to comprehensively evaluate your model.</li><li><strong>Convenience:</strong> Utilize built-in features that remember training settings, simplifying the validation process.</li><li><strong>Flexibility:</strong> Validate your model with the same or different datasets and image sizes.</li><li><a href="https://www.ultralytics.com/glossary/hyperparameter-tuning"><strong>Hyperparameter Tuning</strong></a><strong>:</strong> Use validation metrics to fine-tune your model for better performance.</li></ul><pre>!yolo task=detect mode=val model={HOME}/runs/detect/train/weights/best.pt data={dataset.location}/data.yaml</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*y1XpR7fZOH4eMyTCmaNSCw.png" /><figcaption>Val Results</figcaption></figure><p>General Evaluation:</p><ul><li>Precision, recall, and mAP metrics are quite high.</li><li>Although there are performance differences between classes, the overall model performance is satisfactory.</li><li>It is an effective model in terms of speed, with a low inference time.</li></ul><h4>Running Inference on Test Dataset</h4><p>Let’s evaluate the model’s generalization ability and see its predictions on test datasets it has not seen before by looking at the inferences.</p><pre># predict mode for test data<br>!yolo task=detect mode=predict model={HOME}/runs/detect/train/weights/best.pt conf=0.25 source={dataset.location}/test/images save=True</pre><pre># for visualization<br><br>latest_folder = max(glob.glob(&#39;/content/runs/detect/predict*/&#39;), key=os.path.getmtime)<br>for img in glob.glob(f&#39;{latest_folder}/*.jpg&#39;)[:1]:<br>    display(IPyImage(filename=img, width=600))<br>    print(&quot;\n&quot;)</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/654/1*m6cZFsvjdXDf7EgJIMHgcQ.png" /><figcaption>Result</figcaption></figure><h3>Conclusion</h3><p>I hope you were able to follow along and was able to train successfully.</p><p>I wanted to write this article by gathering new resources for myself and for you, and I hope it has been helpful. Keep following the innovations, and stay updated!</p><p>Endless thanks to the teams at <a href="https://docs.ultralytics.com/"><strong>Ultralytics </strong></a>and <a href="https://roboflow.com/"><strong>Roboflow</strong></a>, pioneers in this field, for their great documentation and open-source contributions.</p><p>If you have any questions, recommendations, or critiques, I can be reached via <a href="https://www.linkedin.com/in/yigitbatuhansener/">LinkedIn </a>or via my <a href="http://ybatu.sener@gmail.com/">mail</a>. Feel free to reach out to me.</p><h4>References:</h4><ul><li><strong>Official Repo: </strong><a href="https://github.com/ultralytics/ultralytics"><strong>https://github.com/ultralytics/ultralytics</strong></a></li><li><a href="https://docs.ultralytics.com/tr/models/yolo11/"><strong>Ultralytics</strong></a></li><li><a href="https://universe.roboflow.com/sixray-dpgc6/weapon_detection-cuahn"><strong>Roboflow</strong></a></li><li><a href="https://www.youtube.com/watch?v=jE_s4tVgPHA&amp;t=606s"><strong>Piotr Skalski’s YouTube broadcast</strong></a></li><li><a href="https://colab.research.google.com/github/roboflow-ai/notebooks/blob/main/notebooks/train-yolo11-object-detection-on-custom-dataset.ipynb"><strong>Orginal Roboflow notebook</strong></a></li><li><a href="https://www.youtube.com/watch?v=rfI5vOo3-_A"><strong>YOLO Vision 2024 Event</strong></a></li></ul><pre>@software{yolo11_ultralytics,<br>  author = {Glenn Jocher and Jing Qiu},<br>  title = {Ultralytics YOLO11},<br>  version = {11.0.0},<br>  year = {2024},<br>  url = {https://github.com/ultralytics/ultralytics},<br>  orcid = {0000-0001-5950-6979, 0000-0002-7603-6750, 0000-0003-3783-7069},<br>  license = {AGPL-3.0}<br>}</pre><pre>@misc{<br>parasites-1s07h_dataset,<br>title = { parasites Dataset },<br>type = { Open Source Dataset },<br>author = { Roboflow 100 },<br>howpublished = { \url{ https://universe.roboflow.com/roboflow-100/parasites-1s07h } },<br>url = { https://universe.roboflow.com/roboflow-100/parasites-1s07h },<br>journal = { Roboflow Universe },<br>publisher = { Roboflow },<br>year = { 2023 },<br>month = { may },<br>note = { visited on 2024-10-01 },<br>}</pre><p>If you enjoy my content, please consider showing your support by clapping for my articles. Your support is greatly appreciated.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2116db587f3b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI-Powered Financial Analysis: Multi-Agent Systems Transform Data into Insights]]></title>
            <link>https://medium.com/@batuhansenerr/ai-powered-financial-analysis-multi-agent-systems-transform-data-into-insights-d94e4867d75d?source=rss-2256a8221220------2</link>
            <guid isPermaLink="false">https://medium.com/p/d94e4867d75d</guid>
            <category><![CDATA[streamlit]]></category>
            <category><![CDATA[multi-agent-systems]]></category>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[stock-market]]></category>
            <dc:creator><![CDATA[Batuhan Sener]]></dc:creator>
            <pubDate>Wed, 31 Jul 2024 00:27:40 GMT</pubDate>
            <atom:updated>2024-07-31T08:15:05.239Z</atom:updated>
            <content:encoded><![CDATA[<h3>AI-Powered Financial Analysis: Multi-Agent LLM Systems Transform Data into Insights</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*L8qMNogL-16IYTKxneJUfQ.png" /></figure><blockquote><a href="https://financial-agent.streamlit.app/">Financial Agents with GPT-4o-mini and LLaMA 3.1 and more … Click here to explore …</a></blockquote><p>This article shares the experiences of developing a multi-agent LLM system that gathers relevant information from the internet, performs sentiment analysis on comments from Reddit, conducts fundamental and technical analysis, and summarizes the information about stock, using a language model. The system is made accessible through Streamlit.</p><h4>Overview of Essential Technologies and Frameworks</h4><p>Before starting the project, I would like to talk about the technologies and frameworks I used in this project:</p><ul><li><a href="https://www.crewai.com/"><strong>1-) CrewAI</strong></a><strong> : </strong>Multi-agent system.</li><li><a href="https://www.langchain.com/"><strong>2-) Langchain</strong></a><strong> : </strong>Tools &amp; model usage.</li><li><strong>3-) </strong><a href="https://www.reddit.com/"><strong>Reddit</strong></a><strong>(praw) &amp; Yahoo Finance(yfinance): </strong>Data collecting.</li><li><a href="https://huggingface.co/models"><strong>4-) Hugging Face</strong></a><strong> :</strong> Finetuned sentiment model for finance.</li><li><strong>5-)</strong><a href="https://groq.com/"><strong> Groq Cloud</strong></a><strong> and</strong><a href="https://platform.openai.com/docs/models"><strong> OpenAI</strong></a><strong> : </strong>Accessing LLama and GPT models.</li><li><a href="https://streamlit.io/"><strong>6-) Streamlit</strong></a><strong> :</strong> UI &amp; Deployment.</li></ul><h4>Explaining Crew AI and How I Used It</h4><p>Crew AI is a collaborative working system designed to enable various artificial intelligence agents to work together as a team, efficiently accomplishing complex tasks.</p><p>Includes crew ai agent tasks and tools.</p><p><strong>Agents </strong>:Each agent is an AI unit that takes on a specific role and works according to that role. Every agent is designed to achieve a specific goal and can perform certain tasks.</p><p><strong>Tasks</strong>: Tasks are specific assignments that agents need to complete, along with instructions on how to complete them. Tasks define the expected outcomes and the tools that will be used.</p><p><strong>Tools</strong>: Tools are additional resources and integrations that agents can use while performing their tasks. These tools provide functions like data collection, analysis, or access to specific APIs.</p><p>In this project, I integrated <strong>Meta </strong>and <strong>OpenAI</strong>’s language models. I created custom tools and developed agents such as Researcher, Technical and Fundamental Analyst, and Reporter to gather stock information. I will share my experiences with this crew system.</p><h4>Explaining Langchain and How I Used It</h4><p><strong>LangChain </strong>is an open-source framework for building applications based on large language models (LLMs). It provides tools and abstractions to customize, improve the accuracy, and relevance of the information generated by the models. Developers can use LangChain components to create new prompt chains or customize existing templates. LangChain also allows LLMs to access new data sets without retraining.</p><p>In this project, I used the YahooFinanceNewsTool for stock information and accessed Groq and OpenAI models.</p><h4>Data Collecting from Reddit and Yahoo Finance</h4><p><strong>PRAW </strong>(Python Reddit API Wrapper) is a library used for easy access to Reddit’s API. PRAW provides a user-friendly interface for fetching Reddit data, browsing subreddits, creating posts and comments, and interacting with Reddit.<br><strong>YFINANCE </strong>is a Python library used to fetch financial data from Yahoo Finance. It allows you to easily obtain various financial data such as stock prices, historical data, financial reports, and other financial information.</p><p>In this project, I used yfinance to collect data for fundamental and technical analysis and for visualization. I used praw to fetch data from Reddit to analyze users’ opinions about the stock from the ‘wallstreetbets’, ‘stocks’, and ‘investing’ subreddits.</p><h4>Explaining Hugging Face and How I Used It</h4><p><strong>Hugging Face</strong> is a company and an open-source platform that provides tools and resources for building, training, and deploying machine learning models. The Hugging Face Hub hosts many models for a <a href="https://huggingface.co/tasks">variety of machine learning tasks</a>.</p><p>In this project, I used the Hugging Face model “<a href="https://huggingface.co/mrm8488/distilroberta-finetuned-financial-news-sentiment-analysis">distilroberta-finetuned-financial-news-sentiment-analysis</a>” to analyze the sentiment of data from Reddit.</p><h4><strong>Groq and OpenAI</strong></h4><p><strong>Groq </strong>is a technology company that develops innovative hardware and software solutions for machine learning and artificial intelligence. Their products are designed to provide high-performance, low-latency computing for complex AI workloads.</p><p>In this project, I integrated GPT-4o, GPT-4o-mini, LLaMA 3 8B, LLaMA 3.1 8B, and LLaMA 3.1 70B language models from<strong> Groq Cloud</strong> and<strong> OpenAI Platform.</strong></p><h4>Streamlit UI and Cloud</h4><p><strong>Streamlit </strong>is an open-source framework used for creating interactive web applications specifically for machine learning, data science and LLM projects. <strong>Streamlit Cloud</strong> is a platform for deploying, managing, and sharing applications directly from GitHub repository.</p><p>In my project, I used Streamlit to create an interactive interface and deployed it on Streamlit Cloud.</p><h4>Steps Covered in this Tutorial</h4><p>The steps to be examined for the multi-agent system:</p><ul><li><strong>Custom Tools for Financial Analysis</strong></li><li><strong>Defining the Crew</strong></li><li><strong>Developing the Streamlit Application &amp; Deployment the App</strong></li><li><strong>Project Usage and Output</strong></li></ul><p>To keep the article concise, I will not delve into the detailed explanation of all the code. For more information, feel free to contact me. My contact details will be shared at the end of the page.</p><h4>1- Custom Tools for Financial Analysis</h4><p><a href="https://docs.crewai.com/core-concepts/Tools/"><strong>Tools</strong></a><strong> </strong>are specialized components that agents can use to perform specific tasks more effectively. These tools can provide various functions such as data collection, analysis, interaction with APIs, or accessing specific resources. You can even integrate it with<a href="https://python.langchain.com/v0.1/docs/integrations/tools/"> L<strong>angchain tools</strong></a>.</p><p>In this article, I will explain the Sentiment tool, but I will also include descriptions of the other tools I used.</p><ul><li><strong>Browser Tool:</strong> Implementing web browsing functionality.</li><li><strong>Search Tools:</strong> Developing specific search operations with <a href="https://serper.dev/"><strong>SERPER</strong></a>.</li><li><strong>Sentiment Analysis Tool:</strong> Building a tool to analyze sentiment in financial data.</li><li><strong>Fundamental Analysis Tool:</strong> Creating a tool for fundamental stock analysis using Yahoo Finance.</li><li><strong>YahooFinanceNewsTool:</strong> Gathering and analyzing stock-related news from Langchain’s tool.</li><li><strong>Tech Analysis Tool:</strong> Implementing technical stock analysis using Yahoo Finance.</li></ul><h4>Sentiment Analysis Tool :</h4><p>First I loads the model and tokenizer from the transformers library for sentiment analysis. The analyze_sentiment function then performs sentiment analysis on given text and returns the result as &quot;negative,&quot; &quot;neutral,&quot; or &quot;positive.&quot; The get_reddit_posts function retrieves posts containing a specific stock symbol from a particular subreddit. Using the praw library, it connects to the Reddit API and fetches the posts. These posts are filtered to include only those from the last 30 days. The ‘ @ tool’ decorator allows the function to be used as a CrewAI tool. The reddit_sentiment_analysis function performs sentiment analysis on posts from specified subreddits about a stock symbol and returns the count of each sentiment label.</p><pre># sentiment_analysis_tool.py<br>import os<br>import praw<br>import torch<br>from datetime import datetime, timedelta<br>from transformers import AutoTokenizer, AutoModelForSequenceClassification<br>from crewai_tools import tool<br><br># Download hf model and tokenizer<br>tokenizer = AutoTokenizer.from_pretrained(&quot;mrm8488/distilroberta-finetuned-financial-news-sentiment-analysis&quot;)<br>model = AutoModelForSequenceClassification.from_pretrained(&quot;mrm8488/distilroberta-finetuned-financial-news-sentiment-analysis&quot;)<br><br>def analyze_sentiment(text):<br>    &quot;&quot;&quot;<br>    Analyze the sentiment of a given text.<br>    &quot;&quot;&quot;<br>    inputs = tokenizer(text, return_tensors=&quot;pt&quot;)<br>    outputs = model(**inputs)<br>    scores = outputs.logits.softmax(dim=1).detach().numpy()[0]<br>    labels = [&quot;negative&quot;, &quot;neutral&quot;, &quot;positive&quot;]<br>    label = labels[scores.argmax()]<br>    return label<br><br>def get_reddit_posts(subreddit_name, stock_symbol, limit=100, days=30):<br>    &quot;&quot;&quot;<br>    Get posts from a specific subreddit containing the stock symbol within the last specified days.<br>    &quot;&quot;&quot;<br>    reddit = praw.Reddit(<br>        client_id=os.getenv(&quot;REDDIT_CLIENT_ID&quot;),<br>        client_secret=os.getenv(&quot;REDDIT_CLIENT_SECRET&quot;),<br>        user_agent=os.getenv(&quot;REDDIT_USER_AGENT&quot;)<br>    )<br>    subreddit = reddit.subreddit(subreddit_name)<br>    end_date = datetime.utcnow()<br>    start_date = end_date - timedelta(days=30)<br>    <br>    posts = []<br>    for post in subreddit.search(stock_symbol, sort=&#39;new&#39;, time_filter=&#39;month&#39;, limit=limit):<br>        post_date = datetime.utcfromtimestamp(post.created_utc)<br>        if start_date &lt;= post_date &lt;= end_date:<br>            posts.append(post.title)<br>    return posts<br><br>@tool<br>def reddit_sentiment_analysis(stock_symbol: str, subreddits: list = [&#39;wallstreetbets&#39;, &#39;stocks&#39;, &#39;investing&#39;], limit: int = 100):<br>    &quot;&quot;&quot;<br>    Perform sentiment analysis on posts from specified subreddits about a stock symbol.<br>    <br>    Args:<br>        stock_symbol (str): The stock symbol to search for.<br>        subreddits (list): List of subreddits to search in.<br>        limit (int): Number of posts to fetch from each subreddit.<br>    <br>    Returns:<br>        list: List of sentiment labels for each post.<br>    &quot;&quot;&quot;<br>    all_sentiments = []<br>    sentiments_counts={&#39;neutral&#39;: 0, &#39;negative&#39;: 0, &#39;positive&#39;: 0}<br>    <br>    for subreddit in subreddits:<br>        posts = get_reddit_posts(subreddit, stock_symbol, limit)<br>        for post in posts:<br>            sentiment = analyze_sentiment(post)<br>            all_sentiments.append((sentiment))<br>            sentiments_counts[sentiment]+=1<br><br>    return sentiments_counts</pre><h4>Defining the Crew</h4><p>For model selection, I created a function called initialize_llm using LLaMA or OpenAI, and wrote the create_crew function to define the tools to be used. I defined Agents and Tasks. When creating agents, I ensured clear goals and detailed background information to make them more capable and specific. For tasks, I provided clear instructions and expected outputs to make them more understandable and achievable. Using the Crew class, I integrated all agents and tasks to form and start the team. Finally, I saved the results to a file and returned them to the user.</p><pre>import os<br>from crewai import Agent, Task, Crew, Process<br>from crewai_tools import SerperDevTool, ScrapeWebsiteTool, WebsiteSearchTool<br>from tools.sentiment_analysis_tool import reddit_sentiment_analysis<br>from tools.yf_tech_analysis_tool import yf_tech_analysis<br>from tools.yf_fundamental_analysis_tool import yf_fundamental_analysis<br>from langchain_community.tools.yahoo_finance_news import YahooFinanceNewsTool<br>from langchain_groq import ChatGroq<br>from langchain_openai import ChatOpenAI<br>from dotenv import load_dotenv<br><br><br><br><br># Environment Variables<br>load_dotenv()<br>os.environ[&quot;SERPER_API_KEY&quot;] = os.getenv(&quot;SERPER_API_KEY&quot;)<br>os.environ[&quot;REDDIT_CLIENT_ID&quot;] = os.getenv(&quot;REDDIT_CLIENT_ID&quot;)<br>os.environ[&quot;REDDIT_CLIENT_SECRET&quot;] = os.getenv(&quot;REDDIT_CLIENT_SECRET&quot;)<br>os.environ[&quot;REDDIT_USER_AGENT&quot;] = os.getenv(&quot;REDDIT_USER_AGENT&quot;)<br><br># Model Selection<br>def initialize_llm(model_option, openai_api_key, groq_api_key):<br>    if model_option == &#39;OpenAI GPT-4o&#39;:<br>        return ChatOpenAI(openai_api_key=openai_api_key, model=&#39;gpt-4o&#39;, temperature=0.1)<br>    elif model_option == &#39;OpenAI GPT-4o Mini&#39;:<br>        return ChatOpenAI(openai_api_key=openai_api_key, model=&#39;gpt-4o-mini&#39;, temperature=0.1)<br>    elif model_option == &#39;Llama 3 8B&#39;:<br>        return ChatGroq(groq_api_key=groq_api_key, model=&#39;llama3-8b-8192&#39;, temperature=0.1)<br>    elif model_option == &#39;Llama 3.1 70B&#39;:<br>        return ChatGroq(groq_api_key=groq_api_key, model=&#39;llama-3.1-70b-versatile&#39;, temperature=0.1)<br>    elif model_option == &#39;Llama 3.1 8B&#39;:<br>        return ChatGroq(groq_api_key=groq_api_key, model=&#39;llama-3.1-8b-instant&#39;, temperature=0.1)<br>    else:<br>        raise ValueError(&quot;Invalid model option selected&quot;)<br><br>def create_crew(stock_symbol, model_option, openai_api_key, groq_api_key):<br>    llm = initialize_llm(model_option, openai_api_key, groq_api_key)<br>    # Tools Initialization<br>    reddit_tool = reddit_sentiment_analysis<br>    serper_tool = SerperDevTool()<br>    yf_tech_tool = yf_tech_analysis<br>    yf_fundamental_tool = yf_fundamental_analysis<br><br>    # Agents Definitions<br>    researcher = Agent(<br>        role=&#39;Senior Stock Market Researcher&#39;,<br>        goal=&#39;Gather and analyze comprehensive data about {stock_symbol}&#39;,<br>        verbose=True,<br>        memory=True,<br>        backstory=&quot;With a Ph.D. in Financial Economics and 15 years of experience in equity research, you&#39;re known for your meticulous data collection and insightful analysis.&quot;,<br>        tools=[reddit_tool, serper_tool, YahooFinanceNewsTool()],<br>        llm=llm<br>    )<br><br>    technical_analyst = Agent(<br>        role=&#39;Expert Technical Analyst&#39;,<br>        goal=&#39;Perform an in-depth technical analysis on {stock_symbol}&#39;,<br>        verbose=True,<br>        memory=True,<br>        backstory=&quot;As a Chartered Market Technician (CMT) with 15 years of experience, you have a keen eye for chart patterns and market trends.&quot;,<br>        tools=[yf_tech_tool],<br>        llm=llm<br>    )<br><br>    fundamental_analyst = Agent(<br>        role=&#39;Senior Fundamental Analyst&#39;,<br>        goal=&#39;Conduct a comprehensive fundamental analysis of {stock_symbol}&#39;,<br>        verbose=True,<br>        memory=True,<br>        backstory=&quot;With a CFA charter and 15 years of experience in value investing, you dissect financial statements and identify key value drivers.&quot;,<br>        tools=[yf_fundamental_tool],<br>        llm=llm<br>    )<br><br>    reporter = Agent(<br>        role=&#39;Chief Investment Strategist&#39;,<br>        goal=&#39;Synthesize all analyses to create a definitive investment report on {stock_symbol}&#39;,<br>        verbose=True,<br>        memory=True,<br>        backstory=&quot;As a seasoned investment strategist with 20 years of experience, you weave complex financial data into compelling investment narratives.&quot;,<br>        tools=[reddit_tool, serper_tool, yf_fundamental_tool, yf_tech_tool, YahooFinanceNewsTool()],<br>        llm=llm<br>    )<br><br>    # Task Definitions<br>    research_task = Task(<br>        description=(<br>            &quot;Conduct research on {stock_symbol}. Your analysis should include:\n&quot;<br>            &quot;1. Current stock price and historical performance (5 years).\n&quot;<br>            &quot;2. Key financial metrics (P/E, EPS growth, revenue growth, margins).\n&quot;<br>            &quot;3. Recent news and press releases (1 month).\n&quot;<br>            &quot;4. Analyst ratings and price targets (min 3 analysts).\n&quot;<br>            &quot;5. Reddit sentiment analysis (100 posts).\n&quot;<br>            &quot;6. Major institutional holders and recent changes.\n&quot;<br>            &quot;7. Competitive landscape and {stock_symbol}&#39;s market share.\n&quot;<br>            &quot;Use reputable financial websites for data.&quot;<br>        ),<br>        expected_output=&#39;A detailed 150-word research report with data sources and brief analysis.&#39;,<br>        agent=researcher<br>    )<br><br>    technical_analysis_task = Task(<br>        description=(<br>            &quot;Perform technical analysis on {stock_symbol}. Include:\n&quot;<br>            &quot;1. 50-day and 200-day moving averages (1 year).\n&quot;<br>            &quot;2. Key support and resistance levels (3 each).\n&quot;<br>            &quot;3. RSI and MACD indicators.\n&quot;<br>            &quot;4. Volume analysis (3 months).\n&quot;<br>            &quot;5. Significant chart patterns (6 months).\n&quot;<br>            &quot;6. Fibonacci retracement levels.\n&quot;<br>            &quot;7. Comparison with sector&#39;s average.\n&quot;<br>            &quot;Use the yf_tech_analysis tool for data.&quot;<br>        ),<br>        expected_output=&#39;A 100-word technical analysis report with buy/sell/hold signals and annotated charts.&#39;,<br>        agent=technical_analyst<br>    )<br><br>    fundamental_analysis_task = Task(<br>        description=(<br>            &quot;Conduct fundamental analysis of {stock_symbol}. Include:\n&quot;<br>            &quot;1. Review last 3 years of financial statements.\n&quot;<br>            &quot;2. Key ratios (P/E, P/B, P/S, PEG, Debt-to-Equity, etc.).\n&quot;<br>            &quot;3. Comparison with main competitors and industry averages.\n&quot;<br>            &quot;4. Revenue and earnings growth trends.\n&quot;<br>            &quot;5. Management effectiveness (ROE, capital allocation).\n&quot;<br>            &quot;6. Competitive advantages and market position.\n&quot;<br>            &quot;7. Growth catalysts and risks (2-3 years).\n&quot;<br>            &quot;8. DCF valuation model with assumptions.\n&quot;<br>            &quot;Use yf_fundamental_analysis tool for data.&quot;<br>        ),<br>        expected_output=&#39;A 100-word fundamental analysis report with buy/hold/sell recommendation and key metrics summary.&#39;,<br>        agent=fundamental_analyst<br>    )<br><br>    report_task = Task(<br>        description=(<br>            &quot;Create an investment report on {stock_symbol}. Include:\n&quot;<br>            &quot;1. Executive Summary: Investment recommendation.\n&quot;<br>            &quot;2. Company Snapshot: Key facts.\n&quot;<br>            &quot;3. Financial Highlights: Top metrics and peer comparison.\n&quot;<br>            &quot;4. Technical Analysis: Key findings.\n&quot;<br>            &quot;5. Fundamental Analysis: Top strengths and concerns.\n&quot;<br>            &quot;6. Risk and Opportunity: Major risk and growth catalyst.\n&quot;<br>            &quot;7. Reddit Sentiment: Key takeaway from sentiment analysis, including the number of positive, negative and neutral comments and total comments.\n&quot;<br>            &quot;8. Investment Thesis: Bull and bear cases.\n&quot;<br>            &quot;9. Price Target: 12-month forecast.\n&quot;<br>        ),<br>        expected_output=&#39;A 600-word investment report with clear sections, key insights.&#39;,<br>        agent=reporter<br>    )<br><br>    # Crew Definition and Kickoff for Result<br>    crew = Crew(<br>        agents=[researcher, technical_analyst, fundamental_analyst, reporter],<br>        tasks=[research_task, technical_analysis_task, fundamental_analysis_task, report_task],<br>        process=Process.sequential,<br>        cache=True<br>    )<br><br>    result = crew.kickoff(inputs={&#39;stock_symbol&#39;: stock_symbol})<br><br>    os.makedirs(&#39;./crew_results&#39;, exist_ok=True)<br>    file_path = f&quot;./crew_results/crew_result_{stock_symbol}.md&quot;<br>    result_str = str(result)<br>    with open(file_path, &#39;w&#39;) as file:<br>        file.write(result_str)<br>    <br>    return file_path</pre><h4>Developing the Streamlit Application &amp; Deployment the App</h4><p>I created an interactive area that allows the user to select a model, integrates API keys for model usage, and generates charts using Yahoo Finance data. The multi-agent system performs all the steps and displays the outputs on the screen. Afterwards, I uploaded the code to GitHub and integrated it with <a href="https://streamlit.io/cloud"><strong>Streamlit Cloud</strong>.</a> To keep this article concise, I am concluding the explanation here. For a more detailed code review, please contact me privately.</p><h4><strong>Project Usage and Output</strong></h4><ul><li>First, select the model you want to use.</li><li>To run the project, you need to obtain an API key based on the model you select. I have shared a link on how you can access <a href="https://console.groq.com/keys"><strong>Groq </strong></a>and <a href="https://platform.openai.com/api-keys"><strong>OpenAI APIs.</strong></a></li><li>Enter the stock symbol you are interested in for a detailed analysis like (NVDA, AAPL, ACN)</li><li>For interactive charts, select the time period and indicators, then click the “Analyze Stock” button.</li></ul><p>The model will then gather data about the selected stock from the internet and provide you with an analysis based on various parameters, including fundamental and technical analysis outputs and sentiment results, according to the template of the chosen model.</p><p><strong>Overview :</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*clcSwhOXl6u9KUhF2MQ8HQ.gif" /></figure><p><strong>Output :</strong></p><pre>Stock Information<br><br>Company Name: NVIDIA Corporation<br><br>Sector: Technology<br><br>Industry: Semiconductors<br><br>Country: United States<br><br>Current Price: $103.73<br><br>Market Cap: $2551581769728<br><br>Analysis Result<br><br>NVIDIA Corporation (NVDA) Investment Report<br><br>1. Executive Summary: Investment Recommendation NVIDIA Corporation (NVDA) is a leading player in the semiconductor industry, particularly known for its graphics processing units (GPUs) that power gaming, AI, and data center applications. Given its robust financial health, strong growth metrics, and competitive positioning, the recommendation is to Hold. While the stock appears oversold, it shows potential for recovery amidst current bearish trends.<br><br>2. Company Snapshot: Key Facts<br><br>Company Name: NVIDIA Corporation<br>Sector: Technology<br>Industry: Semiconductors<br>Market Cap: $2.55 trillion<br>P/E Ratio: 60.66<br>PEG Ratio: 0.89<br>Debt to Equity Ratio: 22.87<br>3. Financial Highlights: Top Metrics and Peer Comparison NVIDIA&#39;s financial performance is characterized by:<br><br>Revenue Growth (YoY): 25.85%<br>Net Income Growth (YoY): 58.13%<br>Gross Margin: 75.29%<br>Operating Margin: 64.93%<br>Net Profit Margin: 53.40% Compared to peers, NVDA&#39;s high P/E ratio indicates strong growth expectations, while its low PEG ratio suggests it may be undervalued relative to its growth potential.<br>4. Technical Analysis: Key Findings<br><br>Current Price: $103.73<br>20-Day Moving Average: $122.04<br>50-Day Moving Average: $119.38<br>RSI: 21.04 (indicating oversold conditions)<br>MACD: Bearish<br>Support Level: $102.54<br>Resistance Level: $136.15 The technical indicators suggest a bearish breakdown, but the stock is currently oversold, indicating a potential for a price rebound.<br>5. Fundamental Analysis: Top Strengths and Concerns Strengths:<br><br>Strong revenue and net income growth rates.<br>Low debt-to-equity ratio, indicating financial stability.<br>High gross and operating margins, reflecting efficient operations.<br>Concerns:<br><br>High P/E ratio may indicate overvaluation.<br>Current bearish trends in the market could impact short-term performance.<br>6. Risk and Opportunity: Major Risk and Growth Catalyst Major Risk: The semiconductor industry is highly cyclical and sensitive to economic downturns, which could adversely affect NVDA&#39;s sales and profitability.<br><br>Growth Catalyst: The increasing demand for AI and machine learning applications presents a significant growth opportunity for NVIDIA, as its GPUs are essential for these technologies.<br><br>7. Reddit Sentiment: Key Takeaway from Sentiment Analysis From the sentiment analysis conducted on Reddit:<br><br>Total Comments Analyzed: 100<br>Positive Comments: 3<br>Negative Comments: 6<br>Neutral Comments: 24 The sentiment is predominantly neutral, with a slight inclination towards negativity, reflecting cautious investor sentiment amidst current market conditions.<br>8. Investment Thesis: Bull and Bear Cases Bull Case: If NVIDIA continues to capitalize on the growing AI market and maintains its competitive edge, it could see substantial revenue growth, leading to a higher stock price.<br><br>Bear Case: Economic headwinds and increased competition in the semiconductor space could lead to declining sales and profitability, negatively impacting the stock price.<br><br>9. Price Target: 12-Month Forecast Considering the current market conditions and NVDA&#39;s growth potential, a conservative 12-month price target is set at $130, reflecting a recovery from current oversold levels and aligning with historical performance trends.<br><br>In conclusion, while NVIDIA Corporation presents a compelling investment opportunity due to its strong fundamentals and growth potential, investors should remain cautious given the current market sentiment and technical indicators. A hold position is recommended until clearer bullish signals emerge.</pre><h4>Conclusion and Recommendations</h4><ul><li>It is not investment advice. Just personal hobby :)</li><li>It’s important to keep the descriptions for Agents and Tasks short and concise. Long and detailed descriptions can exceed the Tokens-per-Minute (TPM) limit and be very costly.</li><li>Before creating the fundamental and technical analysis tools, I was completing tasks by fetching all the data from the internet. This resulted in a high processing load, so I created custom tools to meet this need. This approach is more cost-effective. It’s important to assess needs properly.</li><li>GPT models have shown better inferences compared to LLaMA models. GPT-4o-mini is very functional. It has a higher TPM limit compared to GPT-4, and the results are good. It’s a more cost-effective model.</li></ul><h4>References</h4><p><a href="https://www.reddit.com/"><strong>Reddit</strong></a><strong> (</strong><a href="https://www.reddit.com/r/wallstreetbets/"><strong>wallstreetbets</strong></a><strong>, </strong><a href="https://www.reddit.com/r/stocks/"><strong>stocks</strong></a><strong>, </strong><a href="https://www.reddit.com/r/investing/"><strong>investing</strong></a><strong>)</strong></p><p><a href="https://finance.yahoo.com/"><strong>Yahoo Finance</strong></a></p><p><a href="https://docs.crewai.com/"><strong>CrewAI Docs</strong></a></p><p><a href="https://docs.streamlit.io/"><strong>Streamlit Docs</strong></a></p><p><a href="https://python.langchain.com/v0.1/docs/modules/tools/"><strong>Langchain Tools</strong></a></p><h4><a href="https://financial-agent.streamlit.app/"><strong>You can use the link to try the project</strong></a></h4><p><a href="https://github.com/YBSener/financial_Agent"><strong>All codes are my Github</strong></a></p><p>If you would like to consult or get in touch, you can reach me via <a href="https://www.linkedin.com/in/yigitbatuhansener/"><strong>LinkedIn </strong></a>or <a href="http://ybatu.sener@gmail.com/"><strong>email</strong></a>.</p><p>If you enjoy my content, please consider showing your support by clapping for my articles or buying me a coffee. Your support is greatly appreciated.</p><figure><a href="https://buymeacoffee.com/batuhansenerr"><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Nh1owl6a3HdaRB_Y0y04pw.png" /></a></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d94e4867d75d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[YOLOv10 Custom Object Detection]]></title>
            <link>https://medium.com/@batuhansenerr/yolov10-custom-object-detection-bd7298ddbfd3?source=rss-2256a8221220------2</link>
            <guid isPermaLink="false">https://medium.com/p/bd7298ddbfd3</guid>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[yolov10]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[yolo]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Batuhan Sener]]></dc:creator>
            <pubDate>Sun, 26 May 2024 21:06:35 GMT</pubDate>
            <atom:updated>2024-05-26T21:06:35.916Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Overview of YOLOv10 and Training a Model with Custom Data</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/455/1*T_vXR8hPLhm2uEAv5wEnNg.png" /></figure><h4>Overview</h4><p>YOLOv10, developed using the <a href="https://ultralytics.com/"><strong>Ultralytics</strong> </a>Python package by <strong>Tsinghua University researchers</strong>, offers a novel approach to real-time object detection by improving model architecture and eliminating non-maximum suppression (NMS). These optimizations lead to state-of-the-art performance with lower computational demands. Extensive experiments show that YOLOv10 provides superior accuracy-latency trade-offs across various model scales.</p><p>As those who have read my previous articles know, I have shared various projects using YOLO models because, among pre-trained models, YOLO models stand out significantly in terms of performance and efficiency compared to other models. However, real-time object detection has faced challenges due to reliance on non-maximum suppression (NMS) and architectural inefficiencies. YOLOv10 addresses these issues by eliminating NMS and adopting a design strategy focused on both efficiency and accuracy.</p><h4>Architecture</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/852/1*zolmgF6QKY9cdwDePiT9pA.png" /><figcaption>Graphic from the <a href="https://docs.ultralytics.com/models/yolov10/">Ultralytics</a></figcaption></figure><ol><li><strong>Backbone</strong>: Responsible for feature extraction, the backbone in YOLOv10 uses an enhanced version of CSPNet (Cross Stage Partial Network) to improve gradient flow and reduce computational redundancy.</li><li><strong>Neck</strong>: The neck is designed to aggregate features from different scales and passes them to the head. It includes PAN (Path Aggregation Network) layers for effective multiscale feature fusion.</li><li><strong>One-to-Many Head</strong>: Generates multiple predictions per object during training to provide rich supervisory signals and improve learning accuracy.</li><li><strong>One-to-One Head</strong>: Generates a single best prediction per object during inference to eliminate the need for NMS, thereby reducing latency and improving efficiency.</li></ol><h4>Model Variants and Performance</h4><p><strong>YOLOv10 </strong>comes in six models :</p><ul><li><strong>YOLOv10-N</strong>: Nano version for extremely resource-constrained environments.</li><li><strong>YOLOv10-S</strong>: Small version balancing speed and accuracy.</li><li><strong>YOLOv10-M</strong>: Medium version for general-purpose use.</li><li><strong>YOLOv10-B</strong>: Balanced version with increased width for higher accuracy.</li><li><strong>YOLOv10-L</strong>: Large version for higher accuracy at the cost of increased computational resources.</li><li><strong>YOLOv10-X</strong>: Extra-large version for maximum accuracy and perf</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/555/1*jM9RFJ8H5Pxl5JKMaBHiDg.png" /><figcaption>Graphic from the <a href="https://docs.ultralytics.com/models/yolov10/">Ultralytics</a></figcaption></figure><h4>Comparisons</h4><p>Let’s look at the comparisons between different models regarding latency and accuracy, tested on standard benchmarks like COCO.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/851/1*ycQewXnIe3B4R_u1JEN-Qw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/638/1*z8O_zK-m18l5I1Q2wK_lKg.png" /></figure><p>It is clear that YOLOv10 is the cutting-edge technology for real-time object detection applications, offering higher accuracy and speed performance with fewer parameters.</p><h4>Training YOLOv10 for Custom Object Detection</h4><p>First, clone the official YOLOv10 <a href="https://github.com/THU-MIG/yolov10?tab=readme-ov-file">GitHub </a>repository to download the necessary yolov10n model.</p><pre>!pip install -q git+https://github.com/THU-MIG/yolov10.git<br><br>!wget -P -q https://github.com/jameslahm/yolov10/releases/download/v1.0/yolov10n.pt</pre><p>You can experiment with any custom project on <a href="https://universe.roboflow.com/"><strong>Roboflow Universe</strong></a>, create your own datasets, and even use <a href="https://universe.roboflow.com/roboflow-100"><strong>RF100 </strong></a>datasets sponsored by Intel. For this article, I will use a <a href="https://universe.roboflow.com/vladutc/x-ray-baggage"><strong>pre-prepared dataset</strong></a> designed to detect dangerous items in X-ray images.</p><p>Download your model in YOLOv8 format using the Roboflow API.</p><pre>!pip install -q roboflow<br>from roboflow import Roboflow<br>rf = Roboflow(api_key=&quot;your-api-key&quot;)<br>project = rf.workspace(&quot;vladutc&quot;).project(&quot;x-ray-baggage&quot;)<br>version = project.version(3)<br>dataset = version.download(&quot;yolov8&quot;)</pre><p>Specify the parameters and file paths, then start the model training.</p><pre>!yolo task=detect mode=train epochs=25 batch=32 plots=True \<br>model=&#39;/content/-q/yolov10n.pt&#39; \<br>data=&#39;/content/X-Ray-Baggage-3/data.yaml&#39;</pre><p>Example data.yaml file</p><pre>names:<br>- Gun<br>- Knife<br>- Pliers<br>- Scissors<br>- Wrench<br><br>nc: 5<br><br>roboflow:<br>  license: CC BY 4.0<br>  project: x-ray-baggage<br>  url: https://universe.roboflow.com/vladutc/x-ray-baggage/dataset/3<br>  version: 3<br>  workspace: vladutc<br><br>test: /content/X-Ray-Baggage-3/test/images<br>train: /content/X-Ray-Baggage-3/train/images<br>val: /content/X-Ray-Baggage-3/valid/images</pre><p>Let’s look at the results.</p><pre>Image(filename=&#39;/content/runs/detect/train/results.png&#39;, width=1000)</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/748/1*WfYKEkYSaKjzwIW9mq6fOw.png" /></figure><p>Let’s predict the test data and display the results in a 5x2 grid.</p><pre>from ultralytics import YOLOv10<br><br>model_path = &#39;/content/runs/detect/train/weights/best.pt&#39;<br>model = YOLOv10(model_path)<br>results = model(source=&#39;/content/X-Ray-Baggage-3/test/images&#39;, conf=0.25,save=True)</pre><pre>import glob<br>import matplotlib.pyplot as plt<br>import matplotlib.image as mpimg<br><br>images = glob.glob(&#39;/content/runs/detect/predict/*.jpg&#39;)<br><br>images_to_display = images[:10]<br><br>fig, axes = plt.subplots(2, 5, figsize=(20, 10))<br><br>for i, ax in enumerate(axes.flat):<br>    if i &lt; len(images_to_display):<br>        img = mpimg.imread(images_to_display[i])<br>        ax.imshow(img)<br>        ax.axis(&#39;off&#39;)  <br>    else:<br>        ax.axis(&#39;off&#39;)  <br><br>plt.tight_layout()<br>plt.show()</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/977/1*hjQYhr3XTpsUoKj15-MPug.png" /></figure><h4>Conclusion and Recommendations</h4><ul><li>While creating this article, I trained the YOLOv10n model on multiple datasets, which exhausted my 15GB free T4 GPU limit on Colab. When you exceed the limits while training a model in the Colab environment, there are limitations for the T4 GPU. You can potentially address this issue by logging in with different Google accounts.</li><li>Since technology is advancing rapidly, I believe it is beneficial to learn the main concepts without getting stuck on a single technology in both computer vision and large language models. To adapt to this, it is helpful to learn from the developers of these technologies. The content from <a href="https://github.com/ultralytics"><strong>Ultralytics </strong></a>and <a href="https://github.com/roboflow/notebooks"><strong>Roboflow </strong></a>is very valuable in this area, and it is worthwhile to follow them.</li></ul><h4>References</h4><ul><li><strong>Official Repo</strong>: <a href="https://github.com/THU-MIG/yolov10">https://github.com/THU-MIG/yolov10</a></li><li><a href="https://docs.ultralytics.com/tr/models/yolov10/#key-features"><strong>Ultralytics</strong></a></li><li><a href="https://universe.roboflow.com/sixray-dpgc6/weapon_detection-cuahn"><strong>Roboflow</strong></a></li></ul><pre>@article{THU-MIGyolov10,<br>  title={YOLOv10: Real-Time End-to-End Object Detection},<br>  author={Ao Wang, Hui Chen, Lihao Liu, et al.},<br>  journal={arXiv preprint arXiv:2405.14458},<br>  year={2024},<br>  institution={Tsinghua University},<br>  license = {AGPL-3.0}<br>}</pre><pre>@misc{<br>x-ray-baggage_dataset,<br>title = { X-Ray Baggage Dataset },<br>type = { Open Source Dataset },<br>author = { vladutc },<br>howpublished = { \url{ https://universe.roboflow.com/vladutc/x-ray-baggage } },<br>url = { https://universe.roboflow.com/vladutc/x-ray-baggage },<br>journal = { Roboflow Universe },<br>publisher = { Roboflow },<br>year = { 2024 },<br>month = { may },<br>note = { visited on 2024-05-26 },<br>}</pre><p>I would like to thank the researchers from Tsinghua University, the teams at Ultralytics and Roboflow, as well as all the contributors to the open-source community.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bd7298ddbfd3" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>