<?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 Aram on Medium]]></title>
        <description><![CDATA[Stories by Aram on Medium]]></description>
        <link>https://medium.com/@zerofilter?source=rss-aa22ebd045cd------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*bXeGvqaVa-nLpTd_cpBBnQ.jpeg</url>
            <title>Stories by Aram on Medium</title>
            <link>https://medium.com/@zerofilter?source=rss-aa22ebd045cd------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 15:41:51 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@zerofilter/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[Memories can be a strange thing…]]></title>
            <link>https://zerofilter.medium.com/memories-can-be-a-strange-thing-deef1696b389?source=rss-aa22ebd045cd------2</link>
            <guid isPermaLink="false">https://medium.com/p/deef1696b389</guid>
            <category><![CDATA[neuroscience]]></category>
            <category><![CDATA[emotional-intelligence]]></category>
            <category><![CDATA[memories]]></category>
            <category><![CDATA[psychology]]></category>
            <dc:creator><![CDATA[Aram]]></dc:creator>
            <pubDate>Wed, 26 Aug 2026 11:03:59 GMT</pubDate>
            <atom:updated>2026-08-26T11:03:59.192Z</atom:updated>
            <content:encoded><![CDATA[<p>Being read into broader neuroscience/psychology as a liked curiosity of mine; from interesting science authors like Diane Ackerman (her poetry, essays, naturalist writings are short of amazing) to stories, film, art, etc. I still get surprised, when say, a deep conversation with good friends of mine leads to some amazing metaphors of a phenomenon/concept of which I had not seen in such representation before. Let it be a reminder to explore these spaces with friends, and don’t judge too hard, not all friends are great at articulating their thoughts in a given moment, but life may surprise you, and I hope it does, truly.</p><p>I don’t recall viewing memories, say in the context of a relationship with someone (friendship, family, dating, etc), as points in 3D space, but that each memory has relationships to all other memories/feelings/etc, and its not longer just a set of points in isolation, but some sort of a mesh that takes a strange shape (and likely not a static shape!). Some can view it as a “blob” and there is no shape/structure, but I won’t deny that there can be a shape there or maybe this is just one projection, one dimension of understanding we happen to have onto it.</p><p>I imagine there is a whole fascinating world of analogs and mappings here, gardens, graphs, landscapes, constellations, music, architecture, weaving, maps, etc, stories each exposing something different about that same strange object. A space worth exploring, at least if those kinds of representations make you curious. But I digress, back to the main thought in my hand.</p><p>And when you give memories justice by considering this representation. You realize that, the recollection of memories is not a simple and innocent act.</p><p>Recalling memories can be a slippery slope into tit-for-tat dynamics (think you start “trading” examples with someone, and that’s not very productive), defensive behaviors, emotional manipulation, gaslighting, etc. Not to say examples are bad, but there needs to be some filtering mechanism toward something that is more exemplar and not basic.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jqkCK9B6WqCpwPFZpxTcWg.png" /><figcaption>Some of my doodling on the ideas/concepts floating around.</figcaption></figure><p>And one can make the argument that, you should just stay, so to speak, in the “abstract”, especially when its coming to feelings and emotions. At the end of the day, it is a feeling, and if someone is feeling that way, trying to nail down the events that led to it can become a huge misdirection, say a fixation on the wrong things, overthinking (with the negative connotation), etc.</p><p>In times like this, adages I’ve heard of the like; “To think less and to feel more” or “To say less but to say more” come to the foreground…</p><p>— — — — — — — — — — — — — — — — — — — — -</p><p><strong>&lt;kind of related :)&gt;</strong></p><p>Something I want to start “tagging” to my writings is what I have been consuming at the times of this expansion of thought.</p><p><strong>Music:</strong></p><ul><li><a href="https://youtu.be/ITYVXUvMtHI?si=3DAY8CsANBo8-aLO">Le Temps de l’amour</a></li><li><a href="https://youtu.be/1cS_i-zDdxg?si=ZPoddBSI2I7ZpgTK">Heaven up there</a></li></ul><p><strong>Readings:</strong></p><ul><li>Letters to a young poet.</li><li>A natural history of the 5 senses.</li><li>Some very helpful articles on substack related to the topics at hands.</li></ul><p><strong>Videos:</strong></p><ul><li><a href="https://www.youtube.com/watch?v=-0wCh5wU3yc">a morphine toast</a> (I’ve been periodically re-watching this and a few videos from this channel since 2018, I just absolutely love this videos).</li></ul><p><strong>&lt;/kind of related :)&gt;</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=deef1696b389" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Profiling a Single-GPU Training Bottleneck with NVIDIA Nsight: Achieving Overlap in PyTorch…]]></title>
            <link>https://zerofilter.medium.com/profiling-a-single-gpu-training-bottleneck-with-nvidia-nsight-achieving-overlap-in-pytorch-77c01787e7f7?source=rss-aa22ebd045cd------2</link>
            <guid isPermaLink="false">https://medium.com/p/77c01787e7f7</guid>
            <category><![CDATA[pytorch-dataloader]]></category>
            <category><![CDATA[profiling]]></category>
            <category><![CDATA[performance]]></category>
            <category><![CDATA[pytorch]]></category>
            <category><![CDATA[cuda]]></category>
            <dc:creator><![CDATA[Aram]]></dc:creator>
            <pubDate>Sun, 31 May 2026 11:31:20 GMT</pubDate>
            <atom:updated>2026-05-31T11:31:20.831Z</atom:updated>
            <content:encoded><![CDATA[<h3>Profiling a Single-GPU Training Bottleneck with NVIDIA Nsight: Achieving Overlap in PyTorch DataLoader Workloads.</h3><p>A researcher described the problem in the most natural way possible: the training job had “DataLoader stalls.” That was the main visible symptom. On 1 H100 80GB, with ResNet34 training at a fixed batch size of bs=355(<em>bs was fixed because the learning rate was tuned to the batch size already</em>), the run was sitting at about 1.21 steps/s, and long idle gaps kept appearing between train steps.</p><p>This was a short ad-hoc performance support project I worked on at <a href="https://www.yerevann.com/">YerevaNN</a>. The goal was not to produce a universal recipe for PyTorch performance, and it was not to rewrite half the stack in search of a prettier benchmark. The job was simpler and more exacting than that: work out where the time was actually going, change the smallest thing that mattered, and leave behind something other people could reuse.</p><p>By the end of the investigation, the run had moved from 1.21 to 5.37 steps/s. The useful part was not only the number. It was the path to it. The initial diagnosis pointed at the wrong layer, and the real fix only became obvious once the step timeline was visible in Nsight Systems. We will see a visual of this later in the article.</p><h3>The actual setup and constraints</h3><p>The workload was running on a single H100 in a Slurm-managed cluster, reading data from Weka-backed storage. The software stack already had most of the obvious optimizations in place: pin_memory=True, non_blocking=True , asynchronous host-to-device copies, persistent workers, and torch.compile(mode=&quot;reduce-overhead&quot;) for the steady-state forward path.</p><p>That matters because the “cheap” explanations were already exhausted. The run was not slow because somebody forgot to turn on pinned memory. It was not slow because the code was still in eager mode. It was not slow because the DataLoader was a single-process default setup. The interesting cases begin after those basics are already present (i.e. low-hanging fruit gains).</p><p>There was another constraint that shaped the investigation: the batch size was fixed at 355 because it belonged to a learning-rate experiment. Reducing the batch size may have made the system easier to feed, and it would also have side-stepped the real question. The task here was to make this workload go faster under the conditions/constraints that it actually had to run in.</p><h3>Why “DataLoader stall” was not a diagnosis</h3><p>The phrase “DataLoader stall” was not really wrong. It was just too vague to be a proper diagnosis.</p><p>A slow training loop can lose time in several places that look similar from the outside and behave very differently once you profile them. The workers may be late returning a batch. The pinning thread may be blocking on host allocation. The host-to-device copy may be expensive. The copy may be cheap and still badly placed because it serializes with compute. The model itself may be the bottleneck, and the idle gap you notice may only be the tail of a longer story.</p><p>At a high level, all of these can collapse into the same complaint: the GPU is waiting. That is why I find “DataLoader stall” useful as an observation and not yet useful as an explanation.</p><p>The questions that mattered were somewhat more precise:</p><ul><li>Is the time disappearing inside model compute, or between train_step ranges?</li><li>Is the CPU blocked on pinned-memory allocation?</li><li>Are H2D copies overlapping with compute, or are they queued behind it?</li><li>Are the cluster resource limits even being enforced correctly?</li></ul><p>That last question was more cluster-specific than the rest, and it still mattered. Early in the work, cgroup memory enforcement was not behaving the way the job request implied. Until that is corrected, RAM conclusions are muddy because the system is not actually running inside the envelope you think you are measuring. In other words, part of the performance work was making sure the environment itself was telling the truth.</p><h3>What I tried first, and what it ruled out</h3><p>Before looking for a deeper fix, I verified that the usual input-pipeline basics were already in place: pinned memory, non-blocking transfers, worker parallelism, and enough prefetch depth to avoid a shallow queue. Those are table stakes for this class of workload, and they were not the reason the run was stuck at 1.21 steps/s. The more interesting question was what happened after those basics were already present. That is where the Nsight trace became decisive.</p><h3>What the trace showed</h3><p>Nsight Systems turned the vague complaint about data loading into something more useful, a trace that made the waiting “visible”.</p><p>The long gaps were showing up between train_step ranges, not inside the body of the compiled compute itself. More importantly, the host-to-device copies and the compiled graph kernels were both landing on the same CUDA stream. In this run, that meant stream 7. The GPU would finish a copy, begin compute, hit the next boundary, wait for the next copy, and repeat. The queue could be deeper, the workers could be busier, the memory request could be larger, and none of that would create overlap if copy and compute were still serialized on one stream.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oDndmA0DpxJwCAwgK1rfQw.png" /><figcaption><strong>Before: </strong>Nvidia Nsight Systems profiler output of the Resnet34 Training Run on 1-GPU. Notice the green and orange rows are not overlapping. Before the fix, H2D copies and compiled compute were serialized on the same CUDA stream, so each train step had to wait for the next batch transfer to finish before compute could begin.</figcaption></figure><p>That was the keypoint. DataLoader tuning can help you produce the next batch sooner. It cannot, by itself, make the next batch arrive on the device while the current one is still computing if both operations are ordered on the same stream.</p><p>A compact way to say it is this: the loader was not the whole bottleneck, and the stream schedule was part of the bottleneck. Once that became visible, the fix narrowed sharply.</p><h3>More about the stream-level fix</h3><p>The idea was simple enough to explain in a few lines and subtle enough to require careful synchronisation.</p><p>While batch N is computing on the main stream, issue the host-to-device copy for batch N+1 on a dedicated copy stream. When that copy finishes, record a CUDA event. Then, right before the model consumes batch N+1, make the main stream wait on that event. This preserves correctness because the consumer only proceeds once the transfer is complete, and it creates room for overlap because the copy is no longer queued behind compute on the same stream.</p><p>In article form, the core pattern looks like this:</p><pre>copy_stream = torch.cuda.Stream()<br>next_batch = None<br>next_ready = None<br><br>for batch in loader:<br>    if next_batch is None:<br>        with torch.cuda.stream(copy_stream):<br>            next_batch = move_to_device(batch, non_blocking=True)<br>            next_ready = torch.cuda.Event()<br>            next_ready.record(copy_stream)<br>        continue<br><br>    torch.cuda.current_stream().wait_event(next_ready)<br>    current_batch = next_batch<br><br>    with torch.cuda.stream(copy_stream):<br>        next_batch = move_to_device(batch, non_blocking=True)<br>        next_ready = torch.cuda.Event()<br>        next_ready.record(copy_stream)<br><br>    train_step(current_batch)</pre><p>The code that went into the project was a little more complete than this excerpt. It handled nested batch structures, it was env-gated behind STREAMED_PREFETCH=1, and it was wired so the rest of the training path could remain unchanged when the flag was off (<em>a temporary “modularized” implementation that gives flexibility to “stacking” with potentially other future changes</em>). The essential structure, though, is here in the snippet above: one stream for the current batch&#39;s compute, another stream preparing the next batch, and an event boundary where correctness is enforced.</p><p>That correctness point matters more than the flourish of saying “use another stream.” Without the event wait, the consumer can race ahead and read tensors whose copy has not finished. The optimization works because the concurrency is explicit and the dependency is explicit.</p><p>There was also a compatibility question with torch.compile. In this case, the approach remained viable because the copy staging lived outside the compiled graph boundary. The graph still captured the steady-state compute path, and the stream-prefetch wrapper prepared inputs around it.</p><h3>Before and after</h3><p>Once the stream-prefetch path was in place, the trace changed in the way the theory predicted.</p><p>The H2D copy for the next batch moved onto a separate stream row. The compute stream became visibly more continuous. The idle region between train steps shrank because the system was no longer waiting to begin copying only after the previous compute had already finished.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RP_okzaluPBzM8LuGK7Lug.png" /><figcaption><strong>After: </strong>Once we moved H2D onto a separate copy stream, the next batch transfer overlaps with the current batch’s compute, which shrinks the idle gap between train steps and raises steady-state throughput.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BBlG0TEwlaIVoWxui7Ttdw.png" /><figcaption>A more simplified diagram of the overlapping data movement and compute by putting work into separate streams.</figcaption></figure><p>The measured result was 1.21 -&gt; 5.37 steps/s, roughly a 4.4x increase in throughput, with about 72% of H2D time hidden behind compute. A shorter replicate run landed around 5.0 steps/s, which mattered because it showed this was not a single lucky trace.</p><p>One useful side finding emerged in the follow-up runs. Raising num_workers from 12 to 20 did not improve the system even when CPU and RAM were increased alongside it. Throughput regressed. On this Weka mount, 12 was not a timid starting point. It was more of a local optimum. That is worth noting because it is exactly the sort of thing people tend to misremember six weeks later as &quot;we should probably try more workers.&quot;</p><h3>The final and optimal operating point/configuration</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fAnjm1tKcQ2z_Sw5MAi1jQ.png" /><figcaption>The final set of parameters used for the Single-GPU training of a Resetn34 model in our cluster.</figcaption></figure><h3>Learnings from this performance improvement</h3><p>The first lesson is almost embarrassingly basic and still worth repeating: symptoms arrive in broad language, and good performance work consists in narrowing them until the mechanism is visible. “The DataLoader is slow” and “copy and compute are serialized on the same stream” do not belong to the same stage of understanding. Said more plainly, “The DataLoader is slow” is more of a symptom-level observation. “Copy and compute are serialized on the same stream” is a mechanism-level explanation that we discover after we profiled and analysis the workload.</p><p>The second lesson is that NVTX ranges are not cosmetic. They are what turn a crowded Nsight timeline into an argument. Once the train-step structure was labelled, it became possible to say whether time was disappearing inside the step or between steps. That is a small instrumentation choice with a large effect on how quickly you stop chasing the wrong explanation. One should realize that NVTX ranges are tools to help with instrumentation.</p><p>The third lesson, and more of a minor point but still relevant, is infrastructural. Resource-enforcement bugs distort the meaning of the numbers (<em>if you are using something like Slurm or Kubernetes</em>). If memory caps are not really in force, or if the job is not running inside the limits you think you requested, then “more RAM helped” and “more RAM did not help” are both weaker statements than they appear. Verifying the environment is not preamble to performance work. It is part of performance work.</p><p>And there is a more practical lesson I keep relearning: small, reversible/flexible changes are an excellent shape for risky optimization. The prefetch path was env-gated, easy to A/B switch/test in Slurm, and easy to disable if it caused trouble.</p><h3>Sidebar: Making the learning reusable</h3><p>This small project took about two weeks while I was juggling other work in parallel, and the result was worth more than the life of the original run. A useful optimisation should leave behind a clearer shared model of what was slow, what changed, and when the same idea is likely to help again.</p><p>That is why I cared not only about the code change, and also about how to package the learning. The stream-prefetch wrapper itself was small and fairly portable. The trace pattern was recognizable. The supporting explanation could be turned into a markdown guide, an internal talk, or even an agent-ingestible skill with a pre-flight checklist (<em>think Claude Code or Codex</em>): make sure pin_memory, worker parallelism, non-blocking transfers, and queue depth are already in place before you recommend a side-stream copy path. Going forward, I think a short post in Slack and or the internal Wiki for your organization can be enough (<em>people can just copy-paste that information to their AI agent and see if it can be applied to their codebase(s) for their tasks/workloads</em>). Going a step forward to do the “skill packaging” can be nice but not necessary in my opinion. Then again, this is a new space and worth thinking about when similar cases occur in the future. Maybe building an internal library of AI Agent skills that are useful across your organization is quite valuable and not so time-demanding.</p><p>That packaging work is part of the engineering value for me. A lab benefits when a result becomes easier to reapply correctly and harder to “cargo-cult” into the wrong workload.</p><h3>Why I like this kind of work</h3><p>I like this class of problem because it sits across the stack without becoming abstract. The symptom shows up in throughput numbers. The explanation lives in stream semantics and DMA scheduling. The fix touches a handful of lines and an environment flag. And the work is not quite finished until the next person can understand when the trick is useful and when it is not.</p><p>That makes it satisfying in the old-fashioned sense. The machine is more legible afterwards, and so is the team around it.</p><h3>Sources</h3><ul><li><a href="https://chaimrand.medium.com/optimizing-data-transfer-in-ai-ml-workloads-60df62fe1278">https://chaimrand.medium.com/optimizing-data-transfer-in-ai-ml-workloads-60df62fe1278</a></li><li><a href="https://www.amazon.com/Systems-Performance-Engineering-Optimizing-Inference/dp/B0F47689K8">AI Systems Performance Engineering book by Chris Fregly</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=77c01787e7f7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Is Important When It Comes to the “Inosculation” of AI with Software Engineering?]]></title>
            <link>https://zerofilter.medium.com/what-is-important-when-it-comes-to-to-the-inosculation-of-ai-with-software-engineering-dd1175832fb5?source=rss-aa22ebd045cd------2</link>
            <guid isPermaLink="false">https://medium.com/p/dd1175832fb5</guid>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Aram]]></dc:creator>
            <pubDate>Sat, 23 May 2026 18:26:45 GMT</pubDate>
            <atom:updated>2026-05-25T05:28:46.922Z</atom:updated>
            <content:encoded><![CDATA[<h3>What Is Important When It Comes to the “Inosculation” of AI with Software Engineering? (as of May 2026) — Part 1</h3><p><em>What follows is a synthesis of the current landscape: an unpacking of what has transpired, what is presently unfolding, and what soon may break upon the horizon with regards to </em><strong><em>AI&lt;&gt;Software Engineering</em></strong><em>. In this article, I talk about recent SOTA LLM history, frontier labs incentives/priorities, concepts, tools, and mental models.</em></p><p>Software engineering did not gradually absorb AI. It was abrupter than that. There is little point in rehearsing the arc from the ChatGPT release in late 2022 through the proliferation of LLMs into every corner of the field, many others have talked about that story and will continue to. What matters now is the current snapshot, weighted heavily toward the last six to twelve months. The tools change so fast that even engineers who stepped away from the conversation for even a moment of time came back to find the vocabulary, the workflows, and the reasonable expectations had all shifted. This is not catastrophism. It is simply the tempo of the moment, and it demands a particular kind of calibration before anything else can be said usefully.</p><h3>Current snapshot of SOTA Models by Frontier Labs</h3><h4>(Understanding their priorities/incentives to make sense of what tasks their models will continue to get good at).</h4><p>The first thing worth establishing is where the frontier models actually sit. And not in the abstract, but roughly task by task or groupings of tasks. Harvard coined a useful term for this: the <em>jagged frontier</em>. These models are sharp on some tasks and quietly unreliable on others, sometimes within the same session. The jaggedness is the point. A model that writes a well-structured algorithm can, moments later, confidently misremember an API signature (unfortunate) or lose the thread of its own reasoning mid-task. Knowing where those edges sit has become a genuine professional skill. And it’s something that you develop and polish over time as you interact with and use these models (especially in powerful “harness” like Claude Code or Codex).</p><pre># A tip on a small intervention to deal with annoying &quot;forgetting&quot; or &quot;hallucinations&quot;<br># Of rather important but minute details when in long and &quot;polluted&quot; context windows.<br># Make a command that reads important context on invocation.<br># This way you &quot;inject&quot; that context as a reminder during the part when<br># an LLM needs to generate such commands/text and not forget critical details.<br><br>&lt;a prompt like this to your AI model&gt;<br>- You seem to be forgetting details of &lt;X, Y, Z&gt; throughout the conversation.<br>- Build a simple custom skill/command that can be invoked easily in a turn.<br>- This skill will do something simple like &quot;grep&quot; on a `.md` file.<br>- That `.md` file will contain just enough context to prevent forgetting.<br>&lt;/a prompt like this to your AI model&gt;</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*jvqVLHvvAZqbd2Jb" /><figcaption>Hard to remember source but a popular image on X and Reddit trying to illustrate “jagged frontier”.</figcaption></figure><p>In the last 1–2 years, models are not necessarily exactly uniformly improving. They are improving in specific directions, and those directions are not accidental. The frontier labs have visibly shifted training emphasis toward agentic workflows and programming (<em>this is obvious and doesn&#39;t require much evidence</em>). While the internals are a black box, the reasoning is not hard to follow. Reminds yourself that Reinforcement learning (<em>being used in “new” forms by frontier labs</em>) requires an environment with a reliable verifier: something that can tell the model cleanly whether it succeeded or failed. Code is a near-perfect medium for this. Tests pass or they don’t. A pull request either compiles and satisfies the reviewer’s criteria (at least objective criteria) or it doesn’t. That clean signal makes it possible to structure reward functions and run RL at scale in a way that, say, “write me a persuasive essay” simply cannot match. Creative Writing is an example of a domain that is hard to improve on.</p><p>The result is that multi-turn agentic coding (model planning, calling tools, reading output, self-correcting, and eventually solving X task) got noticeably better over the last six to twelve months. <strong>Around December 2025</strong>, it felt less like incremental improvement and more like a step change. The labs will keep pushing this direction: their revenue is heavily weighted toward engineering customers, the training data is rich and accessible, and the RL loop is clean. You don’t need an internal memo to see where the investment is going.</p><p>And why is the model being so good at programming so useful beyond just “coding”? Look at the next image to understand why.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/914/0*3l9d5nDLxrG25wEy" /><figcaption>Anthropic basically saying “native” tool-calling in LLMs is not token efficient and not as good when it comes to quality/capabilities as the models writing code that performs the tool call itself by announcing this change.</figcaption></figure><p>Does this mean that “native” tool-calling in LLMs is not token efficient and not as good when it comes to quality/capabilities as the models writing code that performs the tool call itself? I definitely think so.</p><p>After all, Anthropic wrote on article on this long ago. Recall this post by Anthropic some time ago <a href="https://t.co/8YCa9siSog">https://anthropic.com/engineering/advanced-tool-use</a>. Using this article to make a “spec”, I recall some people made some implementations of this and posted on the claude subreddit, reporting much more token efficiency and accurate tool calling in their Claude Code workflows. Of course this applies to any SOTA model like Codex as well.</p><p>If all SOTA LLMs converge on this conclusion going forward, I think it’s good. I expect models to get much more token efficient when it comes tool calling (and hopefully more accurate/useful). Writing code is just much succincter and accurate! It just makes sense on this pivot in general.</p><p>One could go further and argue more speculative points here; things like getting models good at programming builds “model IQ” that carries over into other disciplines. That steps into the realm of cognitive science but in the domain of LLMs but is something that could probably be measured and experimented on. I can’t recall right now if such experiments do exist, but I imagine they will also have their limitations. It’s probably too early for anything definitive, but worth watching.</p><p>Another signal worth tracking is how system prompts change over time. Occasionally the frontier labs publish them; and other times, they leak. Either way, reading them carefully is quite revealing, not as a curiosity, but as a form of “archaeology” about the model and how its intended to be driven/steered. A system prompt is a record of what the labs discovered the model needed to be told, which is itself a record of how the model behaved before they told it. Watch a lab’s system prompts evolve across 6 or 12 months and you are, in a roundabout way, watching the model’s failure modes evolve. Simon Willison has been doing this kind of close reading for years, and it repays attention. Look at a <a href="https://simonwillison.net/2026/Apr/18/opus-system-prompt/">recent system prompt change</a> as an example.</p><p>And this next image captures another interesting point to consider about how what we are noticing how the model changes. A subtler observation worth considering, one that cuts against the instinct to treat improvement as straightforwardly good. As frontier LLMs get better, they become <em>worse</em> models of human error. Not human intelligence, but human <em>error</em>. Digest that one for a moment.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*A-a1JXickLGH9BPfjAnVWg.png" /></figure><p>This sort of “gap” between usefulness and predictability seems to be where a lot of unpleasant surprises originate from.</p><h3>With Regards to the “Tooling Layer” of Agents, CLIs, and Harnesses That Wrap the SOTA Models.</h3><p>Generating code is now the cheap part. What separates good output from mediocre output is the harness around the model (<em>tools, prompts, skills, filesystem memory, sub-agents, guardrails, context management, etc.</em>) and workflow you hand it before it writes a single line. This is where most of the practical leverage sits in 2026, and it is also where one could argue that lots of engineers are underinvesting.</p><p>Current tools that are hot right now are the agentic CLIs like Claude Code and Codex CLI. They tend to be first-class in a literal sense that new features land there before the desktop app analogs do. Cursor-style chat interfaces are fine but somewhat overrated in my opinion when put next to these CLIs and the official apps coming directly from the frontier labs. If you have not yet driven a CLI end to end on a real workflow, I think you are missing out and you should try using these harnesses at least a few times.</p><p>MCP servers, plugins, and skills are mature enough now for daily use fortunately. They are roughly interchangeable ways to do the same kind of thing: hand the agent extra tools, prompts and context it would not otherwise have. And one shouldn’t reduce their impact. A very well written prompt that is injected at the right time in a conversation can make a big difference.</p><p>A set of common CLI plugins/”tools” worth knowing specifically:</p><ul><li>Context7 or something similar for live, version-correct library documentation.</li><li>LSP integration for Claude Code in your language of choice (Codex does not have this yet).</li><li>Obra’s “Superpowers” if you haven’t looked at it.</li><li>Both Claude and the Codex also have their own “official” plugin marketplaces, and those are the right place to start rather than random repos people are pushing on GitHub.</li></ul><p>There is definitely security risk in cloning random skill packages from the internet, and it is worth being cautious about what you let run inside your agent’s environment.</p><p>The mental model worth internalizing here is simple: an agent is only as good as the tools and context you give it. An agent is only as good as the tools and context you give it. Wiring together the right MCPs and skills for your stack takes time and genuine experimentation, sort of like a craft that you develop and polish over time. What might serves a Python backend team well can differ from what a TypeScript frontend team needs, and the answer rarely arrives before you get some skin in the game. The official marketplaces are a reasonable place to start. The official marketplaces from Anthropic and Codex are a good place to start though.</p><p>Out of everything I’ve covered in this article, I would say this is probably the highest-leverage habit to build. Most complaints along the lines of “the AI is bad at my codebase” are, underneath, probably some degree of “context problems”. The model is not suddenly less capable it, but rather it does not have what it needs to do the job well. Squeezing the right information out of these models is not a trivial task to do.</p><p>A practical habit here one can develop and maintain is keeping durable context in plain .md files that live in the repo. Files that are human-readable, version-controlled, and findable.</p><p>Hierarchical .md files per directory also work well for larger codebases. The file an agent reads first, CLAUDE.md or its equivalent, is load-bearing infrastructure. One stale or wrong line near the top can quietly steer everything the model produces downstream. Treat it like code, because functionally it is.</p><p>“<strong>Context engineering</strong>” (<em>the new term that captures the nature of this well</em>) is probably the least glamorous part of working with these models, and almost certainly the highest-leverage. The common complaint “the AI is bad at my codebase” is, more often than not, a context problem in disguise. The model has not suddenly become less capable, rather it probably does not have what it needs to do the job well.</p><p>One concrete dimension of this: where information sits inside the context window matters as much as whether it is there at all. Research from Liu et al. (2023), tested across 18 models and nearly 200,000 LLM calls, found a measurable accuracy drop when relevant information lands in the middle of the context rather than near the start or end. The model is not reading your .md files the way you read a document. Position is of context in your prompt is not to be dismissed!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*L3uGnD4pECy7IvP3McwsTQ.png" /><figcaption>Image taken from a presentation in a vide on AI Engineer youtube channel (aiDotEngineer). Skeptical of the 71.5x Token Reduction but nonetheless, there are lots of important learnings that are coming out regarding how to work with these models and CLI model harnesses in the best way.</figcaption></figure><p>Another dimension is regarding how errors compound. Bad context does not produce uniformly mediocre output, it produces a cascade. A model that begins with a flawed understanding of your system will build a flawed plan on top of it, and then generate a large pile of code that faithfully executes that flawed plan. The mistake does not stay small. It ramifies. See the next image from a presentation that captures this phenomenon well.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*o1bZAKsAWZe1F3lUT4n2vA.png" /><figcaption>Visualization for “One stale or wrong line near the top can quietly steer everything the model produces downstream”. Source: <a href="https://www.youtube.com/watch?v=IS_y40zY-hc">https://www.youtube.com/watch?v=IS_y40zY-hc</a></figcaption></figure><p>Another dimension that has existed for a long time in general and is much less technical and more organisational is that tacit knowledge is obviously invisible to the model. And all orgs, teams, etc, have some degree of tribal/tacit knowledge. What lives only in a senior engineer’s head, that could be anything from the reasons a particular abstraction exists, the constraints that shaped an API boundary, the incident that made the team distrust a certain pattern, etc. Writing it down is no longer just good practice for onboarding. It is how you convert institutional memory into something the agent can actually use.</p><p>That is where Part 1 ends , to wrap up, we talked about the models, the harnesses around them, and the context that feeds them. Three layers, each one a precondition for the next. Get the context wrong and the harness cannot save you. Get the harness wrong and the model’s raw capability doesn’t matter. And misread where the models actually sit ( <em>their edges, their failure modes, the directions they are being pushed, etc</em>) and your mental model of the whole thing shaky and out of frame.</p><p>Part 2 will take a bit of a different angle. The question shifts from <em>what are these tools doing</em> to <em>what does this mean for the engineer using them</em>. Which skills compound over time and which atrophy quietly. What judgment still sits with humans and why handing it away is a specific kind of debt. And where, in a field drowning in noise, the actual signal lives.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dd1175832fb5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Claude Code is special for not doing RAG/Vector Search and instead just using tool calling…]]></title>
            <link>https://zerofilter.medium.com/why-claude-code-is-special-for-not-doing-rag-vector-search-agent-search-tool-calling-versus-41b9a6c0f4d9?source=rss-aa22ebd045cd------2</link>
            <guid isPermaLink="false">https://medium.com/p/41b9a6c0f4d9</guid>
            <category><![CDATA[openai-codex]]></category>
            <category><![CDATA[cursor]]></category>
            <category><![CDATA[vector-search]]></category>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[ai-coding-tool]]></category>
            <dc:creator><![CDATA[Aram]]></dc:creator>
            <pubDate>Fri, 06 Feb 2026 18:19:06 GMT</pubDate>
            <atom:updated>2026-04-24T09:32:15.266Z</atom:updated>
            <content:encoded><![CDATA[<h3>Why Claude Code is special for not doing RAG/Vector Search and instead just using tool calling (Grep, etc)</h3><h3>The core question we explore here</h3><p><strong>Why does Claude Code use terminal tools (grep, glob, find) for code discovery instead of the embeddings/vector search/RAG approaches that dominate the ecosystem?</strong></p><p><strong>Short answer from Anthropic (Boris, Latent Space podcast, May 2025):</strong></p><blockquote><em>“We tried very early versions of Claude that actually used RAG… Eventually, we landed on just agentic search as the way to do stuff. And there were two big reasons… </em><strong><em>One is it outperformed everything. By a lot.</em></strong><em> And this was surprising.”</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*m8gf1Q5yQYuOxJxDxypMSw.png" /><figcaption>Image from Official Anthropic Course on Claude Code</figcaption></figure><h3>Anthropic’s Stated Reasons (From Primary Sources)</h3><h3>1. Performance (“It outperformed everything. By a lot.”)</h3><p>From Boris (<a href="https://www.latent.space/p/claude-code">Latent Space podcast, May 2025</a>):</p><blockquote><em>“We tried RAG… we tried a few different kinds of search tools. And eventually, we landed on just agentic search… </em><strong><em>One is it outperformed everything. By a lot.</em></strong><em> And this was surprising.”</em></blockquote><p>When asked about benchmarks:</p><blockquote><em>“This was just vibes, so internal vibes. There’s some internal benchmarks also, but mostly vibes. </em><strong><em>It just felt better.</em></strong><em>”</em></blockquote><p>So we are taking this at face value. They have “internal benchmarks” and it says tool calling performs better than RAG for Claude Code use case programming and programming-adjacent tasks). But I remain a bit skeptical for some nuanced edge cases like large codebases, working across multiple codebases, etc.</p><p>One thing that immediately comes to mind is that it is convenient that the method that uses and generate more tokens performs better (ideal for their business model!). Of course, intuitively, approaches that are generating more tokens have more capacity to be better. Nonetheless, I keep in mind that if token number is going up, accuracy is going up, it is a win-win for them, they make more revenue and their tool utilizes their model in a better way.</p><p>This probably implies that they must have trained on high quality tool-use data, with an emphasis on the programming domain! After all, they say they generally focus in writing, programming, and health-care/life sciences.</p><p>And if you have vast tool-use and agentic flow data in training, it gives way for tool calling to outperform vector search approaches.</p><p>It also means that most likely, Claude Code is built in a simple way, unlike other Agentic programs that have the LLM as a core component. Since Claude Code is not entirely open source, we can speculate a bit here. Instead of building complex orchestration logic, retrieval pipelines, or planning systems on top of the model, the harness can lean on the model itself to “drive” decisions. The core is probably something close to a master while loop:</p><pre>while (tool_call_remaining):<br>  execute_tool_call();<br>  pass_results_to_model();<br>  check_tool_call();<br>  </pre><p>But a quick qualifier about this. I think the loop itself is obviously not what contains complication, but rather the context assembly that happens <em>before</em> each model call (injecting project rules, skill schemas, hook outputs) and the state management that happens <em>after</em> (compaction, session persistence, permission checks). “Context Engineering” is the term the kids use nowadays. But the architectural/directionality bet is clear for them: invest in the model’s ability to drive, not in building a sophisticated “driver”.</p><h3>2. Simplicity &amp; Zero Infrastructure</h3><p>Every RAG system begins with a quiet act of distrust:<em> ‘</em>the model cannot be trusted to find things on its own.’</p><p>From Latent Space Podcast:</p><blockquote><em>“There was this whole indexing step that you have to do for RAG. And there’s a lot of complexity… the code drifts out of sync.”</em></blockquote><p>It constitutes a new label:</p><p><strong><em>The Indexing Tax: </em></strong>Every RAG system pays it eventually: staleness, drift, security exposure, and maintenance entropy.</p><p><strong>Infrastructure burden with RAG:</strong></p><ul><li>Index maintenance</li><li>Embedding service dependency</li><li>Vector database operations</li><li>Staleness management</li></ul><p><strong>Claude Code approach:</strong></p><ul><li>No servers</li><li>No indexes</li><li>Just a CLI</li><li>Works immediately on any codebase</li></ul><h3>3. Security Angle</h3><p>Again from Boris in the <a href="https://www.latent.space/p/claude-code">Latent Space podcast episode</a>.</p><blockquote>And then there’s security issues because this index has to live somewhere. And then, you know, what if that provider gets hacked? And so it’s just a lot of liability for a company to do that. You know, even for our code base, it’s very sensitive. So we’re kind of, we don’t want to upload it to a third-party thing. It could be a first-party thing. But then we still have this out-of-sync issue. And agentic search just sidesteps all of that. So essentially, at the cost of latency and tokens, you now have really awesome search without security downsides.</blockquote><p><strong>Things to consider</strong>: Agentic search keeps more of the codebase local. RAG requires storing representations of your codebase somewhere and usually it is done on the entire codebase. Whereas tool calling/agentic-search is “variable” (It is deciding what files to search and explore).</p><h3>4. Alignment with the “Bitter Lesson” Philosophy</h3><p>Boris says the following on why they don’t build complex architectures:</p><blockquote><em>“Everything is the model. Like, that’s the thing that wins in the end. And it just, as the model gets better, it subsumes everything else.”</em></blockquote><p>Don’t forget Anthropic’s principle: <strong>“Do the simple thing first.”</strong></p><h3>A Token Accumulation Problem</h3><p>Each turn, the context grows:</p><pre>Turn 1: [System] + [User Message]                    ~1,000 tokens<br>Turn 2: [System] + [User] + [Turn 1] + [Tool Result] ~2,500 tokens  <br>Turn 3: [System] + [User] + [T1] + [T2] + [Results]  ~5,000 tokens<br>Turn 4: ...                                          ~8,000 tokens<br>Turn 5: ...                                          ~12,000 tokens</pre><p>Most agent token waste comes from uncontrolled discovery. Claude lists directories, opens files speculatively, greps vaguely, then repeats.</p><h3>Why Autoregressive + Tool Calling Works Despite Token Cost</h3><p><strong>The key insight</strong>: Each turn gives the model information to make a <em>better</em> next decision.</p><pre>RAG (Single Shot):<br>  Query → Retrieve → Done (hope it&#39;s right)</pre><pre>Agentic Search (Iterative):<br>  Query → Result → &quot;Hmm, not quite&quot; → Refined Query → Better Result → &quot;Getting warmer&quot; → ...</pre><p>The model can:</p><ol><li><strong>Course correct</strong> based on what it finds</li><li><strong>Narrow down</strong> from broad to specific</li><li><strong>Follow references</strong> (imports, function calls)</li><li><strong>Recognize patterns</strong> that inform next steps</li></ol><p><strong>This kind of dynamic/adaptive refinement is impossible with single-shot retrieval.</strong></p><h3>Trade-off Analysis Matrix</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*G3IaU5I29rJajd1Tj9klkA.png" /></figure><h3>Direct Comparisons Are Limited</h3><p>Recall from the Latent Space interview with Boris:</p><blockquote><em>Q: “In what benchmark?”<br> A: “This was just vibes, so internal vibes. There’s some internal benchmarks also, but mostly vibes.”</em></blockquote><h3>Related Evidence</h3><p><strong>From Render’s AI Coding Agents Benchmark (2025):</strong></p><blockquote><em>“A point of interest: Cursor utilizes a RAG-like system on the local filesystem to gather more context on the codebase, potentially giving it an advantage.”</em></blockquote><p>Result Cursor (with RAG) vs Claude Code (agentic search):</p><ul><li><strong>Cursor won on feature completeness</strong></li><li>Claude Code comparable on task completion.</li></ul><p><strong>From Stephanie Jarmak’s benchmark work (Jan 2026):</strong></p><blockquote><em>“I’m running ablation studies comparing agents with and without access to Sourcegraph MCPs (including Deep Search)…”</em></blockquote><p>This is exactly the comparison we need, but results are lagging unfortunately.</p><h3>Why You Might Prefer Agentic Search</h3><h3>1. Composability (Also a UNIX Philosophy)</h3><p>Agentic search treats the LLM as a composable unit:</p><pre>tail -f app.log | claude -p &quot;Alert me on anomalies&quot;</pre><p>RAG requires infrastructure that’s harder to compose:</p><pre># Need to set up:<br># - Embedding service<br># - Vector DB<br># - Index maintenance daemon<br># - Query interface<br># Then use it</pre><h3>2. Statelessness</h3><p>Agentic search is <strong>stateless</strong> per session:</p><ul><li>No index to manage</li><li>No staleness to worry about</li><li>Every search is against current reality</li></ul><p>RAG introduces <strong>statefulness</strong>:</p><ul><li>Index is state that must be managed</li><li>State diverges from reality over time</li><li>State must be synchronized</li></ul><h3>3. Model as the “Intelligent” Component</h3><p>Where does the “intelligence” go depending on the approach?</p><ul><li><strong>RAG</strong> pushes intelligence to the retrieval system (embedding quality, chunking strategy, reranking)</li><li><strong>Agentic</strong> search keeps intelligence in the model (decides what to search, how to refine, when to stop)</li></ul><p><strong>Bitter Lesson alignment</strong>: As models improve, the agentic approach gets better “for free.”</p><h3>4. Debuggability</h3><p>With agentic search, you can see <em>exactly</em> what the model did:</p><pre>Turn 1: glob &quot;**/*auth*&quot; → 5 results<br>Turn 2: grep &quot;login&quot; → 12 matches<br>Turn 3: Read auth/login.ts → [contents]</pre><p>With RAG, you see:</p><pre>Query: &quot;where is auth?&quot;<br>Results: [chunk_7, chunk_23, chunk_156]</pre><p>The first is auditable. The second is a black box.</p><h3>5. Resilience to Infrastructure Failures</h3><p>Agentic search fails more gracefully:</p><ul><li>Terminal not available? Can’t work.</li><li>Everything else? Works.</li></ul><p>RAG has more failure modes:</p><ul><li>Embedding service down</li><li>Vector DB unreachable</li><li>Index corrupted</li><li>Chunking mismatch</li><li>Embedding model deprecated</li></ul><h3>Prediction: Will Claude Code Change Its Approach?</h3><p>Based on the evidence:</p><p><strong>I think it is likely to stay with agentic search because:</strong></p><ol><li>It’s working (internal adoption at Anthropic)</li><li>Fits the “bitter lesson” philosophy</li><li>Much less infrastructure burden than RAG/Vector Search</li><li>Security advantages for enterprise</li><li>As models improve, agentic search improves “for free” basically</li></ol><p><strong>Possible enhancements within the paradigm:</strong></p><ol><li>Smarter search tool usage (the Grep vs bash grep issues)</li><li>Better context management (compaction)</li><li>Hybrid approaches (local FTS sidecar)</li><li>Sub-agent parallelization for search</li><li>Hierarchical episodic/periodic memory that partially auto-updates (global level and at project level) that manifests in form of markdown files.</li></ol><p><strong>Unlikely to adopt full RAG because:</strong></p><ol><li>Would require infrastructure</li><li>Would introduce staleness problems</li><li>Would add security concerns</li><li>Goes against their stated philosophy</li></ol><h3>What can you takeaway from this for your workflows?</h3><ol><li><strong>Understand the trade-off clearly</strong>: Tokens vs. Infrastructure</li></ol><ul><li>Agentic: Higher per-query cost, zero setup</li><li>RAG: Lower per-query cost, ongoing maintenance</li></ul><p><strong>2. The “right” choice depends on context</strong>:</p><ul><li>Exploring new codebase → Agentic search (works immediately)</li><li>Large stable codebase, high query volume → Consider hybrid/RAG</li><li>Security-sensitive environment → Agentic search (keeps code local)</li></ul><p><strong>3. The model’s reasoning is the differentiator</strong>:</p><ul><li>RAG “outsources” part of “intelligence work” to retrieval</li><li>Agentic search uses model intelligence for search strategy</li><li>As models improve, agentic approach benefits more</li></ul><p><strong>4. Promising Hybrid approaches exit and are continuing to emerge</strong>:</p><ul><li>SQLite FTS5 + ripgrep (and variants like this out there)</li><li>Code graph tools (but this adds infrastructure)</li><li>The sweet spot may be: light local indexing + model-driven query refinement!</li><li>Check <a href="https://news.ycombinator.com/item?id=46616529">https://news.ycombinator.com/item?id=46616529</a></li></ul><h3>Summary</h3><p>Claude Code’s choice of agentic search over RAG is a <strong>philosophical commitment</strong> to:</p><ol><li>Simplicity over infrastructure</li><li>Model intelligence over retrieval engineering</li><li>Freshness over optimization</li><li>Security over convenience</li></ol><p>The trade-off is <strong>tokens and latency</strong> vs. <strong>infrastructure and staleness</strong>.</p><p>For most coding workflows, especially during exploration and development, agentic search works well. For high-volume queries on large stable codebases, hybrid approaches (like FTS sidecars) can reduce token waste without abandoning the core philosophy. Another great thing to add on top of grep will be LSPs for the languages you work in, this will also reduce token efficiency and improve quality of search within codebases, and basically comes for “free” unlike RAG. Claude Code has already established this, Opencode as well, but not Codex CLI. As of 03–07–26.</p><p><strong>Don’t forget that as models get smarter, agentic search gets better automatically. RAG quality generally depends on continuous engineering effort.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=41b9a6c0f4d9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Agritech Summit 2025: Reflections from attending for 1 day]]></title>
            <link>https://medium.com/climate-tech/agritech-summit-2025-reflections-from-attending-for-1-day-8c20da6415d2?source=rss-aa22ebd045cd------2</link>
            <guid isPermaLink="false">https://medium.com/p/8c20da6415d2</guid>
            <category><![CDATA[agriculture-technology]]></category>
            <category><![CDATA[armenia]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[agritech]]></category>
            <dc:creator><![CDATA[Aram]]></dc:creator>
            <pubDate>Sun, 10 Aug 2025 13:37:41 GMT</pubDate>
            <atom:updated>2025-08-31T16:05:29.401Z</atom:updated>
            <content:encoded><![CDATA[<h3>Seaside Startup Summit 2025:- Agritech Summit: Reflections and a Summary of Day 2</h3><p><strong>Date</strong>: July 24, 2025 <br><strong>Program length</strong>: 3 days (I only attended day 2 and a few selected sessions only during this day)</p><h3>Preface</h3><p>I attended a subset of the Agritech Summit and took comprehensive notes in a few talks. This piece is not a full recap. It focuses on the sessions I documented closely and on themes that stood out for Armenia: inclusion, geospatial data, AI in the field, greenhouse robotics, and post‑harvest value.</p><h3>Executive summary</h3><p>A gap is growing between large farms that can adopt digital tools and older or poorer farmers who risk being left out. Voice‑first advisory, service models over hardware sales, and open geospatial data can close part of that gap. AI helps most when it supports concrete farm decisions and simple workflows. Robotics and better post‑harvest storage look like near‑term wins for Armenia if it is combined with training and data sharing.</p><h3><strong>Opening Keynote: Feeding the future (FAO)</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/768/1*PUqDdUjv2naxhFcZicGeng.jpeg" /><figcaption>A picture of one of the slides in this talk.</figcaption></figure><p><strong>Speaker</strong>: Daniela Di Gianantonio (Head of Digital Agriculture, FAO)<br><strong>Speaker remarks:</strong></p><ul><li>Larger farmers can afford and utilize digital tools; older and poorer farmers are falling behind. Armenia’s farm population skews toward older people.</li><li>Mobile internet adoption is high (lots of people have smartphones and the cost of phone plans with internet is cheap). Also, Armenia has quite decent internet connectivity in rural regions, and add to the fact that Armenia now has Starlink as well (although it is too expensive for a farmer right now). These are some enablers for voice‑first advisory for people who work with their hands and cannot type during the day.</li><li>Drones are probably more useful as <strong>services</strong> than as <strong>products in Armenia</strong>: periodic monitoring, ad‑hoc spraying, and other tasks, sold as outcomes rather than devices.</li><li>FAO Agritech Observatory: an open knowledge repository for digital agriculture in the region (<strong>agritechobservatory.fao.org</strong>).</li><li>FAO Digital Villages Initiative: pilots in Armenia (first sites in Lori and Tavush) that bring basic digital services to rural communities.</li></ul><p><strong>My take</strong>: Farmer‑centric design means starting with tasks and time. Voice over chat for advisory is one example. Another is selling work as a service, not pushing devices or dashboards.</p><h3><strong>AI-Enabled Agriculture</strong>: How and where AI can help now</h3><p><strong>Speakers</strong>: Gevorg Poghosyan, Malte von Bloh, Diana Gurzadyan, Dr. Gevick Davoodi</p><h4>Autonomous AI farming experiments (Technical University of Munich)</h4><p><strong>Speaker</strong>: Malte von Bloh <br><strong>Speaker remarks:</strong></p><ul><li>Project question: Can AI take over parts of crop management? Benchmarks include yield prediction, disease risk, pesticide, and fertilizer timing.</li><li><strong>SmartField</strong> project: ~100k field trials with multiple data partners; inputs include cultivation logs, satellite data, weather, pesticide/fertilizer events, and seed data.</li><li>The aim was to <strong>maintain high yields</strong> while raising resource efficiency and performing close to top‑quartile farms.</li><li>Reality check! <strong>B2Farmer is hard</strong>. End-users are diverse and <strong>margins are thin.</strong></li><li>Many agritech startups fall into the classic startup trap where they focus more on tech than the business case. Data is scarce and messy as well.</li></ul><h4>Building an AI-ready workforce</h4><p><strong>Speaker</strong>: Diana Gurzadyan <br><strong>Speaker remarks</strong>:</p><ul><li>Armenia faces climate stress, skilled labor gaps, market volatility, and limited access to smart tools.</li><li>Common AI areas: automated irrigation, early pest/disease detection, advisory services, and agri‑robotics.</li><li>Examples cited: <strong>DARLI Crop Diagnosis</strong>, <strong>Farmer.chat</strong> (Gen AI assistant; usage numbers presented as speaker remarks), <strong>FAO Digital Services</strong> (weather and crop calendars, livestock, marketplaces, nutrition), and <strong>OneSoil</strong> for remote crop monitoring using years of NDVI signals.</li></ul><p><strong>My take</strong>: AI is useful as decision support that&#39;s tied to timing and dosage, not as a general solver. Start with one crop, one region, and measure outcomes.</p><h3>Digital solutions, data, geospatial, and farm robotics</h3><p><strong>Speakers</strong>: Konstantin Nagornyuk, Khachatur Nikoghosyan, Syuzanna Hovsepyan, Massimo Mercadini</p><h4>Location intelligence for agriculture (Esri / ArcGIS)</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/768/1*FAXqEYAWi1ThAK00PRAY3w.jpeg" /></figure><p><strong>Speaker</strong>: Konstantin Nagornyuk <br><strong>Speaker remarks:</strong></p><ul><li><strong>Cropland masks</strong> underpin monitoring and subsidy audits. They can be built by digitizing imagery, collecting field data, and extracting features from satellite scenes.</li><li>Platform demos showed end‑user training notebooks, AutoML workflows for crop detection, and dashboards for crop analysis.</li><li>Pre‑trained <strong>GeoAI models</strong> exist for objects like solar panels and wind turbines, plus general segmentation (SAM‑style tools).</li><li>Esri <strong>licensing</strong> can be <strong>expensive</strong> for startups. A regional <strong>geoportal</strong> released last year provides free access to some datasets and tools for learning and prototyping.</li></ul><p><strong>My take</strong>: For early pilots, combine open satellite time series with a basic field‑boundary layer and a small set of ground‑truth labels. Keep models simple and iterate.</p><h4>Geetak Robotics: Modular and serviceable in the Greenhouse</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/768/1*a7nig531YCdZ-kNxMQX5Yg.jpeg" /></figure><p>Speaker: Khachatur Nikoghosyan <br>Speaker remarks:</p><ul><li>Started as a robotics school; it has projects where they do <strong>applied robotics for common day-to-day tasks in greenhouses</strong>. The focus is on education being practical, applied and hands-on.</li><li>Built a robot called <strong>GRIG</strong> for tasks like counting and classification (of fruits, plants, etc) using OpenCV, plus sensors for temperature and humidity.</li><li><strong>Thomas‑1, </strong>a<strong> </strong>proof of concept: a cable‑powered platform that runs straight tracks for spraying; avoids battery downtime and weight.</li><li><strong>Bob‑1</strong> proof of concept: manages long cable runs for extended range. The team favors modular blocks: perception, motion, planning, and comms.</li></ul><p><strong>My take</strong>: A cable‑powered platform with simple autonomy is a practical way to deliver value in greenhouses before jumping to full autonomy.</p><h4>Farm-level data strategies (This talk revolved around strawberry production)</h4><p><strong>Speaker</strong>: Syuzanna Hovsepyan <br><strong>Speaker remarks</strong>:</p><ul><li>Shared water‑use figures and argued for <strong>precision irrigation</strong> backed by frequent measurements of water chemistry and plant nutrients.</li><li>Emphasized <strong>light interception</strong> and leaf hydraulics as drivers of plant water dynamics and yield.</li><li>Practice includes monthly water analysis, thermal imaging, and frequent small irrigations based on sensor feedback.</li><li>“<em>Agronomists need to learn to work with data to make wiser and more precise decisions.</em>”</li><li>Overall, I loved how quantitative Syuzanna’s approach was, how much data she collects, how she uses that data, and what it reveals. A very high-quality talk, in my opinion.</li></ul><p><strong>My take</strong>: The combination of water quality data and simple rules for irrigation already beats intuition. Start with a few sensors and clear thresholds.</p><h4>Post-harvest and storage (Marvil Engineering in Italy)</h4><p><strong>Speaker</strong>: Massimo Mercadini <br><strong>Speaker remarks (Limited for this)</strong>:</p><ul><li>Post‑harvest storage is a bottleneck for value capture. The company works on <strong>dynamic controlled atmosphere systems</strong> to extend shelf life without chemical residues.</li><li>South Tyrol context: large apple volumes and a culture of quality‑driven storage.</li></ul><p><strong>My take in brief</strong>: Better storage buys time to sell. For Armenia, even small packing houses can benefit from basic monitoring and alerting. This was quite a complex domain, so I was not fully present and did not take detailed notes. It is quite far out of my area of expertise.</p><h4>Sidebox: Geospatial for agriculture primer</h4><p>Some important variables to take into account for a pilot in Armenia:</p><ul><li><strong>Field boundaries</strong>: a clean polygon layer with IDs. Even partial coverage helps.</li><li><strong>Crop types and seasons</strong>: a minimal label set per field to start (e.g., wheat, barley, tomato, orchard).</li><li><strong>Satellite time series</strong>: a few years of surface reflectance with NDVI and related indices.</li><li><strong>Weather</strong>: gridded temperature, precipitation, and radiation; station data if available.</li><li><strong>Ground truth</strong>: small, well‑curated samples from extension agents or growers.</li><li><strong>APIs and notebooks</strong>: simple access and a starter notebook so others can reproduce results.</li></ul><p>Useful early tasks:</p><ul><li>Cropland mask refresh and change flags.</li><li>Yield proxy maps and input‑timing alerts for one crop.</li><li>Greenhouse counting and quality flags from fixed cameras.</li></ul><h3>Ideas that are ready for Armenia (brief)</h3><ul><li>Voice‑first advisory pilot: Armenian and dialect support; tie to FAO Digital Villages sites; start with two crops and a seasonal calendar.</li><li>Open field‑boundary dataset: publish an initial polygon layer as a public good; invite universities and startups to build on it.</li><li>Greenhouse service runs: counting and spraying as a service using cable‑powered platforms; measuring labor hours saved and yield impact.</li><li>Post‑harvest monitoring kit: low‑cost sensors for small cold rooms; simple dashboards and text alerts; focus on temperature and atmosphere stability.</li><li>Farmer data co‑op: pooled, anonymized metrics to improve forecasting and support better credit and insurance terms.</li></ul><h3>Closing</h3><p>These talks reinforced a few important ideas:</p><ul><li>Start with and don’t lose focus on the use case when building a product.</li><li>Use voice interfaces when building products for users whose hands are busy! Sell services over devices, and let data flow in small, reliable pieces.</li><li>If Armenia can align training, open datasets, and a few practical pilots, the path from talk to field looks much shorter.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8c20da6415d2" width="1" height="1" alt=""><hr><p><a href="https://medium.com/climate-tech/agritech-summit-2025-reflections-from-attending-for-1-day-8c20da6415d2">Agritech Summit 2025: Reflections from attending for 1 day</a> was originally published in <a href="https://medium.com/climate-tech">Climate Tech</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Seaside Startup Summit 2025: Reflections and Notes on the Energy and Entrepreneurship Summit.]]></title>
            <link>https://medium.com/climate-tech/seaside-startup-summit-2025-reflections-and-notes-on-the-energy-and-entrepreneurship-summit-b9a1f52d73f7?source=rss-aa22ebd045cd------2</link>
            <guid isPermaLink="false">https://medium.com/p/b9a1f52d73f7</guid>
            <category><![CDATA[energy]]></category>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[armenia]]></category>
            <category><![CDATA[startupş]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <dc:creator><![CDATA[Aram]]></dc:creator>
            <pubDate>Sun, 10 Aug 2025 12:54:33 GMT</pubDate>
            <atom:updated>2025-08-10T12:54:33.661Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RSuL_u10KnQROW1qerC4_Q.png" /><figcaption>The Energy Transition and Entrepreneurship Summit at Seaside Summit 2025.</figcaption></figure><p><strong>Date:</strong> July 21, 2025<br><strong>Location:</strong> Lake Sevan<br><strong>Program length</strong>: One-day energy program</p><h3>Preface</h3><p>The energy track at Lake Sevan spanned policy, finance, and software into a one-day program. I attended only a few sessions and took comprehensive notes in two. This article covers those talks in depth: industry associations and data access for forecasting, financing a CapEx-heavy sector, and where software can help right now. It is not a full recap of the day, just the sessions I took notes on and followed closely.</p><p>This year’s energy track ran in a single day and packed a lot into two sessions. Below is a write‑up that braids together takeaways from <strong>Fostering Innovation</strong> and a later <strong>Q&amp;A</strong> <strong>on expanding solar companies regionally</strong>. I focus on what seemed actionable for Armenia’s renewable ecosystem, from market structure and associations to forecasting, finance, and software.</p><h3>Compared with 2024</h3><p>In 2024, the energy track ran over two days and ranged wider in topics, I believe. This year, it seemed that the conversation was compressed into a single day at Lake Sevan, with a sharper focus on sector associations, data access for forecasting, and practical finance for a CapEx-heavy industry. The shorter format puts coordination and execution front and center.</p><h3>Session 1: Fostering Innovation</h3><p><strong>Time:</strong> 11:40–13:00<br><strong>Participants</strong>: Moderator <strong>Hayk Shekyan</strong> (Shtigen), <strong>Hakob Hakobyan</strong> (SSS, Startup Armenia), <strong>Karen Asatryan</strong> (R2E2 Fund Director), <strong>Tigran Okhanyan</strong> (EcoVille), <strong>Ronald Upmann</strong> (German Solar Association)</p><h4>German Solar Association (Ronald Upmann): What counts as innovation?</h4><p>Innovation showed up as a mix of technology, process, and market design. The German perspective helped frame it:</p><p><strong>Industry association should be part of the backbone.</strong> Germany’s <strong>Solar Association</strong> is a private‑sector body that spans the value chain, from raw materials to installers and startups. They coordinate policy input, public awareness, and international projects. Armenia lacks a comparable umbrella group with the same scope and capacity.</p><p><strong>Scale and citizen adoption, critical factors.</strong> Panel remarks referenced <strong>17.5 GW</strong> of PV added in Germany last year, with <strong>balcony PV</strong> becoming a mass product. An estimated <strong>500,000 plug‑in balcony kits</strong> have reportedly been sold, signaling how small, low‑friction products can broaden participation.</p><p><strong>Importance of Research -&gt; company pipeline.</strong> Germany’s strong <strong>R&amp;D</strong> footprint, including Fraunhofer institutes, has seeded startup/company spin‑outs. That said, the speaker noted gaps in risk capital and start‑up bureaucracy at home (<em>A perception that even those outside of Germany know about</em>).</p><p><strong>On competing and cooperating globally.</strong> Chinese firms are members of the German association. Rather than pushing tariffs, the association tries to “moderate” among diverse interests and emphasize positive measures for domestic resilience. One tool mentioned was a <strong>resilience bonus</strong>, an added fee or support when buyers choose locally manufactured components like inverters or panels.</p><blockquote>“Balcony PV turned solar into a consumer product.” (Ronald Upmann, from notes)</blockquote><blockquote>“We don’t push for tariffs; many of our members are Chinese manufacturers. Our job is to moderate and keep the market growing.” (Ronald Upmann, from notes)</blockquote><h4>EcoVille (Tigran Okhanyan): new approaches, not only new tech</h4><p>EcoVille pressed for a wider definition of innovation:</p><p><strong>The grid can be more distributed.</strong> Beyond classic central generation, ideas such as buying from batteries that sit off the main grid were raised as examples of market and operational creativity.</p><p><strong>We(Armenia) should upgrade the grid for efficiency and stability.</strong> The call was to identify specific upgrades that boost reliability and enable higher renewable penetration.</p><p><strong>We need to build the workforce for energy.</strong> Armenia needs programs and licenses for hands‑on roles: PV and inverter installers, commissioning techs, and specialized maintenance. Universities and labs will need private support to stand up relevant equipment and projects.</p><p><strong>Match finance to a CapEx‑heavy sector.</strong> Green bonds and tailored lending came up, alongside legislative alignment, so that banks can participate at scale.</p><p><strong>Energy efficiency matters.</strong> Local production of efficiency materials, from windows to roofing, would help reduce demand while supporting industry.</p><blockquote>“Innovation is also how we operate and pay for the grid, not only new hardware.” (Tigran Okhanyan, from notes)</blockquote><blockquote>“The grid doesn’t always have to be centralized.” (Tigran Okhanyan, from notes)</blockquote><h4>R2E2 Fund (Karen Asatryan): Helping round out the ecosystem more</h4><p>R2E2 described itself as a state entity working to create the conditions for renewable energy and efficiency to grow. Points from the panel:</p><p><strong>Convene the full ecosystem.</strong> The system works best when all parties are present and active: firms across the chain, government, financiers, researchers, and civil groups.</p><p><strong>Support local manufacturers.</strong> Armenia already has examples, like <strong>wind‑blade producers</strong>. R2E2 aims to connect them with contracts and event exposure.</p><p><strong>Attract young talent to the field.</strong> The sector needs a pipeline of people who see energy as a career, not a niche.</p><blockquote>“We need the full ecosystem in the room-from manufacturers to financiers-and we can help connect them.” (Karen Asatryan, from notes)</blockquote><blockquote>“Armenia already has wind-blade manufacturing; our job is to open doors to contracts and events.” (Karen Asatryan, from notes)</blockquote><h4>Where software can help right now (Recall the talk is on Fostering Innovation, and naturally, software is an overlapping theme that is top of mind for all people at the summit)</h4><p>Several software opportunities surfaced that can shorten project cycles and improve returns:</p><p><strong>Design and site workflows.</strong> Tools that streamline site surveys, simulation, and installation management.</p><p><strong>Lead generation and qualification.</strong> Better demand capture and customer education reduce friction.</p><p><strong>Performance analytics with learning over time.</strong> Operations tools that track actuals vs. theoretical output and explain deviations.</p><p><strong>Market operations in a liberalized setting.</strong> As wholesale markets open, hundreds of thousands of micro‑decisions appear across trading, scheduling, hedging, and risk. Software and AI can help participants earn a few more cents per kilowatt‑hour by improving bids and balancing.</p><h3>Session 2: Q&amp;A on Regional Expansion for Solar Companies</h3><p><strong>Time:</strong> 15:30–16:30<br><strong>Participants:</strong> Moderator <strong>Hayk Shekyan</strong>; <strong>Tigran Okhanyan</strong> (EcoVille); <strong>Maya Melikidze</strong> (GREDA); <strong>Njdeh Andriazian</strong> (ESF Armenia); <strong>Nick Tabatadze </strong>(Shtigen Georgia)</p><h4>Forecasting and the data problem</h4><p>The strongest theme was forecasting, and the lack of data needed to do it well:</p><p><strong>Engineers cannot access key datasets.</strong> Owners of useful data often do not analyze or share it. In Georgia, the energy exchange data was also cited as hard to open and use.</p><p><strong>Why associations matter.</strong> GREDA was formed because individual companies&#39; requests to the government looked self‑serving and potentially anti-competitive. A “<strong>neutral association</strong>”(<em>but also can be vulnerable to corrupt interests and oligopoly tendencies</em>) can argue for rules and infrastructure that benefit the sector, not just one firm.</p><p><strong>How mature markets handle forecasting.</strong> Some German organizations outsource forecasting to multiple providers in parallel. Armenia largely does not. That reduces redundancy and resilience.</p><blockquote>“When companies lobby alone it looks self-serving; together you can ask for infrastructure that helps the sector.” (Maya Melikidze, GREDA, from notes)</blockquote><blockquote>“We need forecasting providers-plural-and the data access that makes them viable.” (Njdeh Andriazian, ESF Armenia, from notes)</blockquote><h4>Institutions and public policy</h4><p>A recurring frustration was the absence of the “right level” of government in rooms like this. Takeaways:</p><p><strong>Policy often begins outside government.</strong> Civil society and industry can surface problems and shape proposals.</p><p><strong>Define the problem first, in public.</strong> A coalition for sustainable and transparent energy can help set a shared agenda.</p><p><strong>Institutions should hold a mission through policy shifts.</strong> Several speakers said Armenia has policies on paper, but institutions do not consistently carry them through when conditions change.</p><h4>Markets beyond Armenia, but in the Caucasus</h4><blockquote>“Iran has around 8,000 PV specialists today, and they need roughly 28,000 to hit national plans.” (Njdeh Andriazian, from notes)</blockquote><blockquote>Regional ambition came through with specific examples:</blockquote><p><strong>Iran.</strong> Figures cited during the talk suggested <strong>8,000</strong> current PV specialists and a need for <strong>28,000</strong> to meet national goals, alongside incentives that pay a multiple of the regular tariff for renewables. This looks like a skills export and project opportunity if entry barriers can be managed.</p><p><strong>Customer trust in trading.</strong> A new Armenian energy trading company asked how to build trust with customers who are unsure about free market participation. The room leaned toward shared platforms for communication and education.</p><p><strong>Self‑regulation is not enough.</strong> Several argued that the market will not reach fairness or transparency on its own. Multiple bodies already have mandates in the energy sector; they need to coordinate rather than sit in silos.</p><h3>What to do next: a practical shortlist</h3><ol><li><strong>Stand up a renewables industry association with teeth.</strong> Give it a permanent council and clear sphere/scope: policy proposals, data access advocacy, training standards, and international partnerships.</li><li><strong>Launch an open data and forecasting program.</strong> Publish a staged roadmap for data access; seed two to three independent forecasting providers and compare them in public. Use standardized APIs.</li><li><strong>Create installer and O&amp;M credentials.</strong> Co‑design licenses and curricula with universities and private firms. Fund shared labs and field gear.</li><li><strong>Catalytic finance for CapEx.</strong> Encourage green bonds, guarantee facilities, and standardized contracts so banks can underwrite rooftop and C&amp;I projects at scale.</li><li><strong>Prepare for cross‑border trade.</strong> Align grid upgrades, interconnect rules, and forecasting with neighbors to sell skills and services, not just electrons.</li></ol><h3>Open questions worth pushing forward…</h3><ul><li>Which Armenian organization is best placed to host a sector‑wide association, and how should leadership be structured to avoid capture</li><li>What are the minimum viable datasets that unlock useful forecasting in year one, and who can publish them safely?</li><li>For a customer‑facing performance checker, what inputs should be user‑editable and what can be inferred from device telemetry?</li><li>Which grid upgrades yield the largest near‑term stability gains for distributed PV integration?</li><li>What is a realistic path to certify 1,000+ installers and O&amp;M techs in two years, and who funds the labs and exams?</li><li>How can a resilience bonus be designed to reward local value without breaching trade rules or inflating costs?</li></ul><h3>In Closing</h3><p>Attending a few sessions at the Energy summit made it clear how much progress depends on <strong>coordination</strong>. The technology is here. The missing pieces are institutions that share data, train people, set fair rules, and give companies the room to compete. If Armenia can line those up, innovation will look less like a slogan/buzzword and more like day‑to‑day practice.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b9a1f52d73f7" width="1" height="1" alt=""><hr><p><a href="https://medium.com/climate-tech/seaside-startup-summit-2025-reflections-and-notes-on-the-energy-and-entrepreneurship-summit-b9a1f52d73f7">Seaside Startup Summit 2025: Reflections and Notes on the Energy and Entrepreneurship Summit.</a> was originally published in <a href="https://medium.com/climate-tech">Climate Tech</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The different factors that contribute to humans' intelligence and their interesting relationships.]]></title>
            <link>https://zerofilter.medium.com/the-different-factors-that-contribute-to-humans-intelligence-and-their-interesting-relationships-4fcdfa975c5a?source=rss-aa22ebd045cd------2</link>
            <guid isPermaLink="false">https://medium.com/p/4fcdfa975c5a</guid>
            <category><![CDATA[iq]]></category>
            <category><![CDATA[intelligence]]></category>
            <category><![CDATA[neuroscience]]></category>
            <category><![CDATA[cognitive-science]]></category>
            <dc:creator><![CDATA[Aram]]></dc:creator>
            <pubDate>Wed, 05 Feb 2025 10:59:38 GMT</pubDate>
            <atom:updated>2025-02-05T10:59:38.702Z</atom:updated>
            <content:encoded><![CDATA[<h3>The different factors that contribute to humans&#39; intelligence and their interesting relationships. An Apriori algorithm-like analysis.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kpNXrFLGzMD7tGazUGn5aQ.png" /></figure><p>How do you think these variables interact dynamically over a lifetime? Think of something like an Apriori Algorithm run on these variables and brain outcomes.</p><h3>Dynamic Interactions Over a Lifetime: An Apriori-Inspired Analysis</h3><p>The Apriori Algorithm identifies frequent itemsets (variable combinations) that predict outcomes. Applying this logic to cognitive development, we might uncover associative rules linking genetic, neural, and environmental factors to outcomes like prodigy abilities or high IQ. Here’s how these variables might interact dynamically:</p><h4>1. Early Childhood (Ages 0–10): Synaptic Pruning &amp; Environmental Priming</h4><p><strong>Key Variables:</strong><br>- Genetic predisposition<br>- Synaptic pruning efficiency<br>- Early environmental stimulation</p><p><strong>Associative Rules:</strong><br>- Genetic predisposition + enriched environment → accelerated synaptic pruning in domain-specific networks (e.g., auditory cortex in music prodigies).<br>- Reduced synaptic pruning (autism-linked) + high working memory → hyper-specialized skills (e.g., calendar calculation in savants).</p><p><strong>Outcomes:</strong><br>- Prodigies emerge when domain-specific neural circuits (e.g., visual-spatial for math, auditory for music) are amplified by early exposure and practice.<br>- High IQ trajectories begin with efficient hippocampal-prefrontal connectivity, enhanced by activities demanding working memory (e.g., puzzles, language learning).</p><h4>2. Adolescence (Ages 10–25): Neural Efficiency &amp; Specialization</h4><p><strong>Key Variables:</strong><br>- White matter integrity<br>- Neural efficiency<br>- Deliberate practice</p><p><strong>Associative Rules:</strong><br>- High working memory + domain-specific practice → strengthened corticostriatal loops (skill automation).<br>- Autism-like attention to detail + environmental support → “prodigy-to-polymath” transitions (e.g., von Neumann’s shift from math to economics).</p><p><strong>Outcomes:</strong><br>- Neural efficiency consolidates: High IQ individuals optimize task-related networks (e.g., prefrontal cortex for logic), while prodigies deepen domain mastery.<br>- Critical period closing: Synaptic pruning tapers, making late-stage skill acquisition harder without prior scaffolding.</p><h4>3. Adulthood (25+): Plasticity, Compensation, &amp; Legacy</h4><p><strong>Key Variables:</strong><br>- Cognitive plasticity<br>- Environmental complexity<br>- Compensatory strategies</p><p><strong>Associative Rules:</strong><br>- High neural efficiency + continued learning → neurogenesis in hippocampus (maintaining episodic memory).<br>- Reduced synaptic pruning (autism) → trade-offs: expertise retention vs. rigidity.</p><p><strong>Outcomes:</strong><br>- Polymaths leverage cross-domain plasticity (e.g., Einstein’s physics-to-philosophy insights).<br>- Prodigies risk stagnation without environmental novelty but may transition to teaching/mentoring roles.</p><h3>Beyond Math, Music, and Art: Other Prodigy Categories</h3><p>While math, music, and art are well-documented, prodigies exist in other domains, though they’re less studied:</p><h4>1. Linguistic Prodigies</h4><p>- Master multiple languages or invent new grammars by age 10 (e.g., “hyperpolyglots”).<br>- Likely tied to an enhanced phonological loop (working memory) and Broca’s area plasticity.</p><h4>2. Chess/Strategic Prodigies</h4><p>- Excel in pattern recognition and probabilistic forecasting (e.g., Magnus Carlsen).<br>- Overlap with math prodigies in visual-spatial and working memory strengths.</p><h4>3. Athletic Prodigies</h4><p>- Exceptional motor memory and cerebellar coordination (e.g., young gymnasts or soccer phenoms).<br>- May involve procedural memory dominance over declarative memory.</p><h4>4. Social/Emotional Prodigies</h4><p>- Rare but possible; e.g., children with precocious empathy or leadership skills (linked to mirror neuron system efficiency).</p><h4>5. Technology-Related Prodigies? (Potential Derivative of Math/Logic/Science)</h4><p>- Coding or engineering savants (e.g., early mastery of programming logic), potentially tied to fluid reasoning and frontoparietal network efficiency.</p><h3>Unanswered Questions &amp; Future Directions</h3><ol><li>Feedback Loops: How do early successes (e.g., a prodigy winning competitions) alter environmental inputs (e.g., parental investment), creating self-reinforcing cycles?</li><li>Neural Trade-offs: Does hyperconnectivity in autism or prodigies come at the cost of metabolic efficiency or resilience to aging?</li><li>Cultural Biases: Are some prodigy domains overlooked due to societal values (e.g., chess vs. farming)?</li><li>Genetic-Environmental Thresholds: What minimum thresholds of genetic predisposition and environmental enrichment are needed to trigger prodigy-level abilities?</li></ol><h3>Final Synthesis</h3><p>Exceptional cognition arises from nonlinear interactions between genetic predispositions, neural wiring, and environmental catalysts. Over a lifetime, these variables form dynamic feedback loops:</p><p>- Prodigies: Domain-specific circuits + early practice + reduced pruning → rapid expertise.<br>- High IQ: Efficient networks + broad curiosity → adaptive problem-solving.<br>- Autism Spectrum: Local hyperconnectivity + environmental scaffolding → deep but narrow mastery.</p><p><strong>The “Apriori rules” of genius might resemble:<br></strong>- IF high working memory + enriched environment + DRD4–7R gene?→ THEN prodigy potential.<br>- IF efficient pruning + high plasticity + intellectual stimulation → THEN polymath trajectory.</p><p>Ultimately, the brain’s multidimensional landscape allows genius to emerge through diverse pathways — whether via efficiency, hyperconnectivity, or obsessive focus. New prodigy categories will likely emerge as we expand our understanding of human potential.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4fcdfa975c5a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Noisy Neighbor Problem in Multi-Tenant Systems Explained Briefly]]></title>
            <link>https://zerofilter.medium.com/noisy-neighbor-problem-in-multi-tenant-systems-explained-briefly-3788ae5e9d5b?source=rss-aa22ebd045cd------2</link>
            <guid isPermaLink="false">https://medium.com/p/3788ae5e9d5b</guid>
            <category><![CDATA[multi-tenant]]></category>
            <category><![CDATA[noisy-neighbors]]></category>
            <category><![CDATA[distributed-systems]]></category>
            <category><![CDATA[multi-tenant-architecture]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Aram]]></dc:creator>
            <pubDate>Tue, 04 Feb 2025 21:38:46 GMT</pubDate>
            <atom:updated>2025-02-04T21:38:46.115Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*soRUkNLjA77_ndBbPUfVlw.png" /><figcaption>Breaking solutions down by Cost and Isolation for Noisy Neighbor problem in Multi-Tenant Systems.</figcaption></figure><p>Multi-tenant systems power many modern cloud and SaaS platforms, but sharing resources comes at a price. Just like in a shared apartment where one roommate hogs the remote or the kitchen, one tenant’s excessive resource use can negatively affect everyone else. In this article, we’ll explore the noisy neighbor problem — from its basic definitions to the challenges emerging in AI/ML workloads — and examine strategies to mitigate its impact, drawing on anonymized insights from real-world industrial platforms.</p><h3><strong>Intro: Resource Sharing Problem</strong></h3><p>Imagine living with several roommates where one person constantly uses <strong>more than their fair share of shared resources</strong>. In multi-tenant systems, a similar dilemma arises: when one tenant consumes disproportionate resources, it degrades the performance of others. This “<strong>noisy neighbor</strong>” issue is not only a technical headache — it directly affects user experience, business reliability, service level agreement guarantees offered to customers, scalability, and more. Today’s systems aim for a balance between efficient resource sharing and adequate isolation.</p><h3><strong>1. Understanding Multi-Tenant Systems</strong></h3><p>Multi-tenancy evolved from single-tenant setups to maximize cost efficiency and scalability by sharing hardware and software resources across multiple customers. However, this evolution brings a fundamental tension:</p><h4><strong>Cost Efficiency vs. Isolation:</strong></h4><p>Shared resources lower costs for the application/system but increase the risk of one tenant’s heavy usage impacting others.</p><ul><li><strong>Trade-Offs: </strong>Architects must balance performance, security, and cost, ensuring that the system can handle varying loads while protecting against resource contention.</li></ul><h4><strong>Anatomy of the Noisy Neighbor Problem</strong></h4><p>At its core, the noisy neighbor problem occurs when one tenant’s activities overload shared resources — whether that’s compute cycles, database capacity, network bandwidth, or message processing systems. Consider these common cases:</p><ul><li><strong>Compute Resources: </strong>A tenant running intensive operations can starve others of processing power.</li><li><strong>Database Contention: </strong>High-volume queries from one tenant may slow down overall system performance.</li><li><strong>Network and Messaging: </strong>Shared message brokers or queues (e.g., those using ordering keys) can experience blockages if one tenant’s workload lags, delaying others.</li></ul><h4><strong>Real-World Impact: Anonymized Case Studies</strong></h4><ul><li><strong>Industrial IoT Platform: </strong>One anonymized example comes from an industrial AI platform ingesting data from tens of thousands of sensors. When one industrial facility (a tenant) generated a burst of data, it threatened to delay or block data processing for other refineries.</li><li><strong>AI Infrastructure Example: </strong>In a recent discussion at an anonymized startup, developers noted that shared messaging systems sometimes lead to one tenant’s heavy processing (or even accidental misfires like a script running in an unintended loop) blocking the delivery of messages for other customers. These experiences underline the critical need for smart resource management.</li></ul><h4><strong>Implications in AI/ML proliferation: New Dimensions of the Problem</strong></h4><p>Modern AI and ML workloads add further complexity:</p><ul><li><strong>Diverse Workloads: </strong>From batch processing in ensemble ML systems to continuous data streams in real-time data streams, the range of resource requirements is vast.</li><li><strong>Unpredictable Demands: </strong>Workloads like RL Agent inference/training, fine-tuning, and large language model (LLM) inference can be sporadic and intense, leading to sudden spikes in resource usage.</li><li><strong>Mixed Task Management: </strong>Systems that handle both deep learning tasks and real-time operations must dynamically balance competing needs, further complicating resource isolation.</li></ul><h3><strong>2. Architectural Patterns for Mitigation</strong></h3><p>Effective strategies for mitigating the noisy neighbor problem involve both isolating resources and managing workloads smartly.</p><h4><strong>Resource Isolation Strategies</strong></h4><ul><li><strong>Shared Nothing Architecture: </strong>One proven approach is to allocate dedicated resources per tenant. For example, an anonymized insight from someone at a startup in the industrial AI space mentioned that “we would’ve had a noise problem with our current database design, but that could have been resolved by making a separate database for each tenant.” While this <strong>database-per-tenant approach</strong> can greatly reduce interference, it comes with <strong>increased costs.</strong></li><li><strong>Dedicated Clusters or Resource Pools: </strong>For high-demand or critical tenants, setting up separate clusters or isolated resource pools ensures that one tenant’s workload doesn’t impact others.</li></ul><h4><strong>Workload Management Patterns</strong></h4><ul><li><strong>Queue and Message Processing Designs: </strong>Implementing ordering keys and segregated processing queues can help ensure that a slow or heavy workload from one tenant doesn’t block others.</li><li><strong>Rate Limiting and Throttling: </strong>Algorithms like the token bucket can cap a tenant’s request rate, preventing any one tenant from monopolizing shared resources.</li><li><strong>Traffic Sidelining: </strong>Redirecting misbehaving or overly aggressive traffic into separate processing channels allows for continued service for all tenants without interruption.</li></ul><h4><strong>Database Strategies</strong></h4><ul><li><strong>Tenant Isolation Approaches: </strong>Depending on the system’s needs, architects might opt for a multi-tenant database with robust resource management or a database-per-tenant model that offers greater isolation at a higher cost.</li></ul><h3><strong>3. Emerging Challenges in the AI Era</strong></h3><p>As AI and ML workloads become more integrated with multi-tenant systems, new challenges emerge:</p><ul><li><strong>Real-Time ML Inference: </strong>Low-latency requirements demand that systems not only handle bursts in data but also process critical tasks without delay.</li><li><strong>LLM Deployments: </strong>Large language models require intensive computation, and isolating these operations from other workloads is crucial to maintain overall performance.</li><li><strong>Edge Computing and Hybrid Workloads: </strong>As more processing moves to the edge, balancing the load between centralized and distributed systems adds another layer of complexity to resource management.</li></ul><h3><strong>4. Conclusion: Designing for Coexistence</strong></h3><p>The noisy neighbor problem highlights a key challenge in modern multi-tenant architectures: designing systems that allow for both cost-efficient resource sharing and effective isolation. By understanding the root causes — from compute and database contention to unpredictable AI/ML workloads — and by employing strategies such as shared-nothing architectures, rate limiting, and workload segregation, system architects can better prepare for these challenges.</p><p>Real-world insights, like those of anonymized discussions from startups (<em>from personal experience and conversations with fellow software engineers</em>), remind us that while the technical solutions may come at a cost, proactive planning and smart design choices are essential. As workloads evolve and new challenges emerge, designing for coexistence remains a cornerstone of building resilient, scalable multi-tenant systems.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3788ae5e9d5b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Can’t Context Length Just Be Bigger for AI Models Like ChatGPT?]]></title>
            <link>https://zerofilter.medium.com/why-cant-context-length-just-be-bigger-for-ai-models-like-chatgpt-a75fd7181a17?source=rss-aa22ebd045cd------2</link>
            <guid isPermaLink="false">https://medium.com/p/a75fd7181a17</guid>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[prompt-engineering]]></category>
            <category><![CDATA[in-context-learning]]></category>
            <category><![CDATA[retrieval-augmented-gen]]></category>
            <category><![CDATA[content-length]]></category>
            <dc:creator><![CDATA[Aram]]></dc:creator>
            <pubDate>Mon, 03 Feb 2025 10:20:06 GMT</pubDate>
            <atom:updated>2025-02-03T10:20:06.295Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*r8-3U7hVUjRjWJDyw-cCaA.png" /><figcaption>Brief overview of pros/cons of simply increasing context length.</figcaption></figure><p>Modern AI models like ChatGPT capture our imagination with their ability to hold extended conversations and process large chunks of text. At first glance, it might seem that simply allowing these models to consider even more context should boost their performance. However, as explored in recent discussions — and as highlighted in <em>AI Engineering</em> — the reality is more nuanced.</p><h3><strong>Context Length scales quadratically, this is the real curse.</strong></h3><p>Transformer models process information through an attention mechanism that computes relationships between every pair of tokens in an input. In practice, this means the computational cost scales quadratically with the sequence length (O(n²)). In simple terms, doubling the number of tokens doesn’t just double the work — it quadruples it. This “attention bottleneck” poses a serious challenge when considering longer context windows.</p><p>For example, as explained in a recent Medium article on context windows, adding more tokens increases the number of key and value vectors the model must process, which quickly ramps up both latency and compute costs cite .</p><h3><strong>Efficiency Versus Capability</strong></h3><p>While our intuition suggests that more context equals better performance, there is a delicate balance at play:</p><p>• <strong>Compute Costs:</strong> Longer sequences demand significantly more processing power. This isn’t merely a theoretical concern; in real-world applications, the extra tokens can slow down response times and inflate energy costs.</p><p>• <strong>Diminishing Returns:</strong> Empirical studies on scaling laws for language models indicate that additional context yields logarithmically smaller benefits even as costs increase exponentially cite .</p><p>• <strong>Risk of Dilution:</strong> There’s a greater chance that the model might focus on less relevant parts of the input, potentially degrading the quality of the output. In other words, simply adding more tokens can sometimes lead to more noise than signal.</p><h3><strong>Rethinking Context: Quality Over Quantity</strong></h3><p>Instead of a brute-force expansion of the context window, the focus is shifting toward smarter methods of handling context. One promising approach is Retrieval-Augmented Generation (RAG), where models are designed to fetch and incorporate only the most relevant pieces of information from a larger corpus. This method sidesteps the need to process an enormous amount of text at once while still benefiting from a vast information base.</p><p>Other innovative techniques involve hybrid architectures that combine efficient retrieval with robust transformer-based processing. These models balance the need for a moderate context window with improved attention mechanisms that use the available tokens more effectively.</p><h3><strong>The Future: Aim for Smarter, Not Just Longer Context</strong></h3><p>The key takeaway is that raw context length isn’t the sole determinant of an AI model’s performance. Instead, how effectively a model can utilize its available context is what truly counts. A hybrid strategy that leverages both efficient context retrieval and optimized attention mechanisms may be the real breakthrough.</p><p>In summary, while extending context windows — as seen in models like Gemini — is an advantage, it’s only one piece of a complex puzzle. The future of AI model development lies not in endlessly stretching the context window, but in finding smarter ways to construct and utilize context that balance efficiency, quality, and cost.</p><p>This nuanced approach is a reminder that when it comes to AI, bigger isn’t always better — it’s about being intelligently bigger.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a75fd7181a17" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Truffle-1: Your Personal Lego Block of AGI]]></title>
            <link>https://zerofilter.medium.com/truffle-1-your-personal-lego-block-of-agi-8e4806116c4f?source=rss-aa22ebd045cd------2</link>
            <guid isPermaLink="false">https://medium.com/p/8e4806116c4f</guid>
            <category><![CDATA[decentralization]]></category>
            <category><![CDATA[truffle-1]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[agents]]></category>
            <dc:creator><![CDATA[Aram]]></dc:creator>
            <pubDate>Sun, 26 Jan 2025 15:54:47 GMT</pubDate>
            <atom:updated>2025-01-27T09:33:47.456Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*WbuoyYn8DLcjQQ3t.jpg" /><figcaption>This but with orbs that light up (Original Image from the anime Psycho Pass)</figcaption></figure><p><strong>tl;dr;<br>- </strong>Truffle-1 flips the script on recent AGI developments and visions (think a decentralized network of specialized geniuses vs one mega-brain)<br>- A custom Mixture of Models running on a separate device that can connect to your laptop. Basically imagine the potential to do unlimited local AI requests and or access a network of other truffles for certain tasks.<br>- Runs on an Nvidia Jetson AGX Orin (64GB powerhouse in a slick glowing case). The hardware can be upgraded easily later, they can just swap the SOC (looking at recent Nvidia DIGITS and future innovations to be done by hardware providers). Imagine the potential future differentiation of truffle-1 base and truffle-1 pro (akin to MacBook pro products).<br>- Brings agent-based computing to your desk with infinite local inference. (think about that for a few seconds)<br>- Creates a marketplace of AI capabilities (agents trading knowledge and offering services like a decentralized brain economy).</p><h3>From “AI in a datacenter” to “AI in your backpack”</h3><p>Let’s talk about something wild: while everyone’s chasing AGI by building bigger models in massive datacenters, Truffle-1 is taking a completely different path. It’s not just another AI product — it’s reimagining how we scale intelligence itself to achieve superintelligence.</p><p>Here’s the thesis: superintelligence won’t emerge from some gigantic single model. Instead, it’ll come from networks of specialized agents working together, just like how human civilization created a kind of collective superintelligence through collaboration and specialization.</p><p>To really get why this matters, check out this contrast:<br>- Dario Amodei’s vision (OpenAI alum, co-founder of Anthropic(Claude)): <a href="https://darioamodei.com/machines-of-loving-grace">“A Country of geniuses in a datacenter”</a><br>- Truffle-1’s take: “A decentralized network of custom, diverse and varied agents/geniuses working together on tasks”</p><h3>The Brain in the Box (but make it digital)</h3><p>According to their <a href="https://x.com/0xLienid/status/1869450375004381448">system card (as of December 2024)</a>, Truffle-1’s architecture is fascinating stuff. They’ve basically built a digital version of specialized brain regions:</p><p>1. <strong>Audiocortex</strong>: Your AI’s ears (processing speech and sound)<br>2. <strong>Speechcortex</strong>: Your AI’s voice box (making thoughts audible)<br>3. <strong>Neocortex</strong>: The thinking machine (where the magic happens)</p><p>This isn’t just bio-mimicry for the sake of it — it’s about efficient resource use. Just like your brain doesn’t activate everything at once, Truffle-1 routes tasks to specialized models (they call it MoM — Mixture of Models).</p><h3>“Chat is dead, long live tools” (the dev perspective)</h3><p>Here’s where it gets a bit more interesting (and controversial maybe?). Truffle-1 completely ditches the chat paradigm we’re all used to. Instead, everything is about tools and tasks. Maybe chat was what we needed for our intro to these models and how to interact/interface with them but it is not the most optimal way.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dr88lQzeIjsDrUihqy4eBw.png" /><figcaption>Embedding the X post was not letting me publish the article.</figcaption></figure><p>Their SDK docs show how apps are just Python files full of tools (think: OpenAI Gym but for real-world stuff). Check out their sample app “Yang” — instead of chatting about trading strategies, it’s out there actually analyzing markets and making moves.</p><p>An example of a function annotated from their <a href="https://github.com/deepshard/trufflesdk/blob/main/hisnameisyang.py">1st example on the Git Hub</a>.</p><pre># you may provide the tool description for the model and icon (right now Apple SF Symbol names) here, as well as indicate that this function should be treated as a tool<br>    @truffle.tool(<br>        description=&quot;Returns metadata given a path to a CSV file, such as column names, data types, and shape&quot;, <br>        icon=&quot;tablecells.badge.ellipsis&quot;<br>    )<br>    @truffle.args(path=&quot;The path to the CSV file to analyze&quot;) # we went back and forth with ways to provide these arg descriptions, definitely would love feedback<br>    def AnalyzeCSV(self, path: str) -&gt; Dict[str,Any]:<br>        if not os.path.exists(path):<br>            return {&quot;error&quot;: f&quot;File &#39;{path}&#39; not found&quot;}<br>        try:<br>            with open(path, &#39;r&#39;) as file:<br>                df = pd.read_csv(path)<br>                return {<br>                    &quot;path&quot;: path,<br>                    &quot;columns&quot;: df.columns.tolist(),<br>                    &quot;shape&quot;: df.shape,<br>                    &quot;dtypes&quot;: df.dtypes.apply(lambda x: x.name).to_dict()<br>                }<br>        except Exception as e:<br>            return truffle.ReportError(e) # returns detailed traceback and error info for the model to troubleshoot withp</pre><p>Some key things that make this work (from their technical docs):<br>- Tools are tokenized and added directly to model vocabulary.<br>- Apps run in sandboxed containers (safety first in case AI models go crazy and try to mess up your computer).<br>- Everything is about action loops, not chat loops (big stuff here, think in terms of tools, functions, actions, etc).</p><h3>The Network is the Computer (but this time for real)</h3><p>Remember Sun’s old slogan? Truffle-1 actually makes it real. Their Agent Registry Protocol (<a href="https://github.com/deepshard/AgentRegistry">detailed in their January 2025 docs</a>) creates this fascinating marketplace of AI capabilities.</p><p>Real-world example: They’re running this pilot in Nigeria where local clinics use Truffle-1s to connect with specialized medical agents. It’s like having a team of international doctors on call, but at local prices. That’s the power of decentralized AI (and it’s just the beginning).</p><h3>Relevant Economics insights and interesting ideas (because incentives matter)</h3><p>This is where it gets crypto-interesting (but in a good way). Their protocol includes:<br>- Agent registration on-chain (with hardware-backed security).<br>- Token rewards for useful AI work.<br>- Staking mechanics for quality control.<br>- Reputation systems that actually matter.</p><h3>Why This Could Be Huge (beyond the hype)</h3><p>Looking at their multi-model inference benchmarks <a href="https://x.com/dylanvsthe/status/1869451506375856449">(released January 2025)</a>, Truffle-1 is showing some serious performance:<br>- Parallel processing of multiple models on edge hardware<br>- Competitive performance with much larger cloud-based systems<br>- Real innovation in resource management (not just marketing fluff)</p><h3>The Road Ahead?</h3><p>Think about the evolution of personal computing:<br>- 1970s: Your own computer (revolutionary!)<br>- 1980s: Graphical interfaces (game-changer)<br>- 1990s: The internet (mind-blown)<br>- 2020s: Your personal AI network? (that’s the Truffle-1 bet)</p><p>Is this the future? Maybe. Is it a fascinating alternative to the “bigger is better” approach? Absolutely in my opinion. This is why I have pre-ordered the 1st batch of Truffle-1s and can’t wait to get mine. Every day I spend some time pondering what agents I can develop on it, how I can use it, etc.</p><p>What’s clear is that we’re seeing a real attempt to solve the multi-agent scaling problem that everyone’s worried about (see their approach to agent coordination and knowledge management in their system card). Whether it works or not, the lessons learned here will be invaluable.</p><p><em>This piece connects dots from Truffle-1’s public docs, their system card (December 2024), protocol specs (January 2025), and ongoing pilot programs. All specs and capabilities are current as of early 2025. Let me know if you want to dive deeper into any part!</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8e4806116c4f" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>