The End of the $20/month AI Subscription.

We need to talk about what this version means for all those pricey monthly subscriptions we keep paying.

Let’s look at GLM-5.2.

This is a massive open-weight language model. Zhipu AI launched it in June 2026. The specs are ridiculous. It has a one million token context window. It handles project-level coding tasks. It also manages complex reasoning. You do not have to pay a hefty monthly fee to a closed ecosystem to get this kind of power. You can just grab the weights off Hugging Face. It has an MIT license. You can also use a metered API for cheap.

Many people are comparing it to top-tier closed models. It holds its own. It actually matches or beats heavily funded subscription models on coding benchmarks. The model has specific reasoning modes. You can set it to high for everyday tasks. Or you can crank it to max for profound problem solving.

Then there is HappyHorse 1.1.

Alibaba built this video generator. It handles text-to-video and image-to-video. It even extends existing clips. It generates cinematic footage with native audio. The lip sync is built right in. The motion is stable.

Here is the kicker. You access it through platforms like WaveSpeedAI on a pay-per-call basis. It costs about 70 cents a generation. No monthly minimums. No recurring subscriptions.

This is a huge problem for the big paid models, which usually charge twenty or thirty bucks a month just for access to a decent LLM.

We are used to paying twenty or thirty bucks a month just for access to a decent LLM. We pay another fee for a video generator. We get locked into their ecosystems. But these new releases are taking a different route. They are putting out frontier-level tech for free or on a strict pay-as-you-go basis.

Why pay a heavy monthly fee when you can self-host GLM-5.2 for complex coding? Why stay locked in a video tool subscription when HappyHorse 1.1 lets you pay only when you actually render a video?

These releases force the market to rethink pricing. Open weights give developers control. Metered APIs save money. The whole industry is tired of endless subscriptions anyway.

The old model of hoarding access behind a monthly paywall is starting to crack. The tools are getting better. The access is getting cheaper. That is a massive win for everyone trying to build something new.

Let’s talk about the specs behind GLM-5.2 first.
It runs on a massive 1-million-token context window. That means you can dump an entire software repository into the prompt. It handles long-horizon coding workflows and autonomous agent loops without dropping the context. Zhipu AI built it with two distinct reasoning tracks. You use “High” mode for routine coding tasks. If an agent hits a wall, it kicks into “Max” mode for profound problem solving. This is the kind of granular control you usually only see in premium enterprise tiers. GLM-5.2 gives it away under a permissive MIT license.

Paid LLM subscriptions strictly limit your token usage. If you run a heavy agent loop on a closed model, you hit rate limits incredibly fast. Or you rack up a massive API bill. GLM-5.2 lets you run these massive operations on your hardware. If you cannot self-host, you just pay for the compute required. You do not pay a premium just to access the model weights.

Now look at HappyHorse 1.1.
Alibaba engineered this model specifically for cinematic visual quality. It outputs crisp 1080p footage for up to 15 seconds. You get text-to-video and image-to-video generation right out of the box. There is also a highly capable reference-to-video mode. This lets you upload a character image and generate entirely new scenes while keeping that person visually consistent. The model even executes multi-shot sequencing from a single prompt. Describe the scene, and it figures out the camera movement and cut structure on its own.

Subscription video generators usually lock you into a rigid monthly credit system. You pay thirty bucks just for the right to try. Then your credits vanish on failed generations, forcing you to upgrade. HappyHorse 1.1 operates entirely differently on platforms like WaveSpeedAI. You access the model through a single REST API. You pay exactly 70 cents per generation. No platform fees. No recurring subscriptions. No minimum spend.

This pricing shift changes how we work. You stop worrying about burning through a monthly allowance. You just pay for the exact compute you use. When open-weight LLMs and metered video APIs match the output of premium subscriptions, the old business model fractures. Developers get total control. Creators get affordable access. The days of hoarding frontier tech behind a rigid paywall are numbered.

In Summary:
We are watching the ground shift under the subscription model. You no longer have to rent your tools by the month just to get decent results. Models like GLM-5.2 and HappyHorse 1.1 prove that high-end AI does not have to mean high-end recurring fees. You can run massive coding workflows or generate cinematic shots on your own terms. Pay for what you use. Keep the rest of your money. It is a much better way to build.

Last week, my coding agent started answering me as if it were a ghost.

There was no press release, no Silicon Valley keynote with dramatic lighting, and no founder doing victory laps on X. Just a new slug quietly added to OpenRouter labeled stealth/ox-alpha with a price tag of exactly zero dollars.

Within forty-eight hours, half the developer group chats I am in were buzzing with excitement.

People were feeding entire legacy codebases into it, throwing wild Three.js rendering bugs at it, and watching it solve problems that usually choke top-tier proprietary models. The wildest part? Nobody officially knew who made it.

Now that the dust is settling, here is what Ox Alpha actually is, why the mystery captured everyone’s attention, and whether you should really care.

What Is Ox Alpha?

On paper, Ox Alpha looked like a typo or a private staging build someone accidentally pushed to production.

It dropped with a massive 1,048,576-token context window, a 131,072-token maximum output, and multimodal support for text, images, and video. For a preview model, the infrastructure behind it was absurd: reports circulated that the mysterious provider had provisioned capacity for around 100 trillion tokens a day.

You can access it right now on OpenRouter, making it trivial to drop into your CLI, Cursor, or any agent framework running on OpenAI-compatible endpoints.

Naturally, developers immediately tried to break it.

Why the AI Community Turned into Internet Sleuths

The internet loves a compelling mystery, and Ox Alpha was catnip for tech forensics.

Within days, researchers used every method they could think of to unmask the creator. They analyzed video token consumption rates (which scale at an oddly specific 147 tokens per second), tested prompt wrappers, and looked at tokenization quirks.

The forensic trail pointed heavily toward Beijing: specifically Zhipu AI (Z.ai) and the GLM model lineage. Shortly after the buzz peaked, reports confirmed what the fingerprints had already revealed: Ox Alpha is the stealth testing ground for the latest iteration of the GLM flagship series.

Releasing models under anonymous handles on router platforms has quietly become a favorite playbook for AI labs. This approach allows them to stress-test infrastructure at a massive scale and gather unvarnished, blind-test feedback without brand bias or marketing hype.

How Good Is It in Practice?

Benchmarks are great, but developer workflows are reality.

• Front-End and UI Sense: This aspect is where Ox Alpha genuinely surprised people. Most LLMs can write syntactically correct React or HTML, but the layouts often come out looking like a generic template from 2018. Ox Alpha seems to have a genuine knack for visual spacing, restrained palette choices, and clean typography.
• Long-Context Code Archaeology: Having a million tokens means you don’t need to surgically slice out files before asking a question. Feeding multi-file repositories to trace race conditions across workers and locks worked remarkably well without dropping variable names midway through.
• Speed and Quirks: It is not flawless. At around 35 to 40 tokens per second, it is not the fastest runner on the track. It also inherits standard LLM habits: it rarely writes unit tests unless you explicitly hold its hand and demand them.

The Reality Check: Watch Your Data

Free frontier-level compute is intoxicating, but there are always caveats.

When Ox Alpha first surfaced anonymously on OpenRouter, the third-party provider's data retention policies stated that it stored prompts and completions. Even with no training usage claimed, it’s never a good idea to feed proprietary intellectual property, API keys, or private customer records to an anonymous endpoint.

The smart move: use it heavily for open-source maintenance, rapid prototyping, and UI design sandboxes. Keep client secrets and private repos behind your standard, verified enterprise providers.

The Takeaway

Ox Alpha proved two things. First, the gap between open-weight frontier reasoning models and expensive closed-source APIs is shrinking at a breakneck pace. Second, the developer ecosystem is now flexible enough that an unannounced, anonymous model can appear on OpenRouter on a Thursday and be running inside thousands of terminal workflows by Friday morning.

If you haven’t plugged your API key into it yet, give it a spin on a tricky bug while the free preview lasts. It might just surprise you.

Aloha! 🌺

Today, we are introducing Ornith-1.0, a self-improving family of open-source models specially for agentic coding tasks. Ornith-1.0 spans the full spectrum, from compact 9B Dense models suitable for edge device deployment to 397B MoE frontier-scale models optimized for maximum performance, with variants including 9B Dense, 31B Dense, 35B MoE, and 397B MoE. Built on top of pretrained Gemma 4 and Qwen 3.5, it achieves state-of-the-art performance among open-source models of comparable size on coding benchmarks.

The key innovation behind Ornith-1.0 is a self-improving training framework. Instead of relying on human-designed harnesses to drive solution generation in RL, Ornith-1.0 learns to generate both solution rollouts and the task-specific harnesses that guide those rollouts. By jointly optimizing the scaffold and the resulting solution, the model can discover better search trajectories and generate higher-quality solutions.

Ornith-1.0 achieves state-of-the-art performance among open-source models of comparable size across a broad range of agentic coding benchmarks: Ornith-1.0–397B (77.5 on Terminal-Bench 2.1 and 82.4 on SWE-Bench Verified) matches the performance of Claude Opus 4.7 (70.3 on TB-2.1 and 80.8 on SWE-Bench Verified) and outperforms leading open-source models of similar size, including MiniMax M3 (66.0 on TB-2.1 and 80.5 on SWE-Bench Verified) and DeepSeek-V4-Pro (67.9 on TB-2.1 and 80.6 on SWE-Bench Verified). Ornith-1.0–9B, which can be easily deployed on edge devices, matches or exceeds the performance of much larger models such as Gemma 4–31B and Qwen 3.6 35B.

At the flagship scale, Ornith-1.0–397B achieves 77.5 on Terminal-Bench 2.1 and 82.4 on SWE-Bench Verified, surpassing Claude Opus 4.7 on both benchmarks and outperforming leading open-source models of similar size, including Minimax M3 and DeepSeek-V4-Pro.

Ornith-1.0–35B significantly outperforms similarly sized models, including Qwen 3.5–35B, Qwen 3.6–35B, and Gemma 31B. Despite having only 35B parameters, it even surpasses Qwen 3.5–397B on Terminal-Bench 2.1 (64.4 vs. 53.5) while matching its performance across several other coding and agentic benchmarks.

The edge-deployable Ornith-1.0–9B also delivers remarkably strong results, achieving 43.1 on Terminal-Bench 2.1 and 69.4 on SWE-Bench Verified. Despite being a compact 9B-parameter model, it matches or exceeds the performance of much larger models such as Gemma 4–31B, demonstrating that strong agentic coding capabilities can be achieved even in resource-efficient deployments.

A Self-improving Strategy for LLM Training

At the core of Ornith-1.0 is a self-improving training framework that jointly learns to solve tasks and to construct the scaffolds that guide those solutions. Rather than relying on a fixed, human-designed harness shared across a task category, Ornith-1.0 treats the scaffold as a learnable object that co-evolves with the policy.

Each RL step proceeds in two stages: conditioned on a task and the scaffold previously used for it, the model first proposes a refined scaffold; conditioned on that scaffold and the task description, it then generates a solution rollout. Reward from the rollout is propagated to both stages, so the model is optimized not only to produce better answers but also to author the orchestration that elicits them.

Repeated over training, this yields a feedback loop in which scaffolds are continually mutated and selected toward those that induce higher-reward trajectories, allowing per-task-category strategies to emerge automatically and driving sustained capability gains without hand-engineered harness design.

Addressing Reward Hacking in Self-Improvement

Allowing the model to author its own scaffold naturally introduces the reward-hacking issue. A self-generated scaffold can learn to satisfy the verifier without performing the task: reading the visible test files and hardcoding the expected artifacts, such as touching the checked-for file or writing the literal expected output, or copying an oracle solution present in the environment.

We defend against this in three layers. First, we fix the outer trust boundary: the environment, the tool surface, and test isolation are immutable and outside the model’s reach, so the model evolves only the inner policy scaffold: its memory, error-handling, and orchestration logic.

Second, a deterministic monitor enforces that boundary at the level it can be specified exactly, flagging any attempt to read withheld paths, modify verification scripts, or invoke actions outside the sanctioned tool surface, and assigning such trajectories zero reward with exclusion from the advantage computation.

Third, because intent-level gaming can occur entirely within the allowed tool surface, a frozen LLM judge acts as a veto on top of the verifier rather than the primary reward.

Asynchronous RL Training

For RL training, to address the off-line policy problem for long rollouts, Ornith-1.0 adopts the pipeline-RL strategy. To control the effect of earlier generated off-policy tokens, we apply a staleness weight that downweights tokens according to their age and drops them entirely once a threshold is exceeded: You can try Ornith here: https://huggingface.co/spaces/ornith-ai/Ornith-1.0-9B

The token-level GRPO loss is weighted as follows:

Full Table

BenchmarkOrnith-1.0–397BQwen3.5–397BQwen3.7-MaxGLM-5.2–744BMinimax-M3–428BDeep Seek-V4-Pro-1.6TClaude Opus 4.7Claude Opus 4.8Terminal Bench 2.1 (Terminus-2)77.553.573.581.0646470.385 Terminal Bench 2.1 (Claude Code) 78.248.669.882.7–66.569.778.9 SWE-Bench Verified 82.476.480.4–80.680.887.6 SWE-Bench Pro 62.251.660.662.15955.464.369.2 SWE-Bench Multilingual 78.969.378.3–76.2—NL2Repo48.236.847.248.942.1–69.7 ClawEval 77.170.765.2–75.878.2-SWE Atlas—QnA41.220.4–37.927.240.348.8SWE Atlas—RF42.618.4—48.646. 7SWE Atlas—TW39.118.5–30.8–38.5-

BenchmarkOrnith-1.0–35BQwen3.5–35BQwen3.6–35 BGemma4–31BQwen3.5–397BTerminal Bench 2.1 (Terminus-2) 64.241.452.542.153.5 Terminal Bench 2.1 (Claude Code) 62.838.949.2–48.6 SWE-Bench Verified 75.67073.45276.4 SWE-Bench Pro 50.444.649.535.751.6 SWE-Bench Multilingual 69.360.367.251.769.3NL2Repo34.620.529.415.536.8 ClawEval Avg: 69.865.468.748.570. 7SWE Atlas—QnA37.113.215.5–20.4SWE Atlas—RF29.710.211.4–18.4 SWE Atlas—TW27.89.813.3–18.5

BenchmarkOrnith-1.0–9BQwen3.5–9BQwen3.5–35 BGemma4–12BGemma4–31BTerminal Bench 2.1 (Terminus-2)43.121.341.42142.1 Terminal Bench 2.1 (Claude Code) 40.618.938.9—SWE-Bench Verified 69.453.27044.252 SWE-Bench Pro 42.931.344.627.635. 7SWE-Bench Multilingual 5239.760.332.551.7NL2Repo27.216.220.510.315.5ClawEval Avg 63.153.265.432.548. 5SWE Atlas—QnA17.99.213.2—SWE Atlas—RF16.64.310.2—SWE Atlas— TW15.34.49.8 —

Footnote

• Terminal-Bench 2.1 (Terminus-2): We evaluate Terminal-Bench 2.1 using the Harbor/Terminus-2 framework with parser=json, temperature=1.0, top_p=1.0, and a 128K context window. Each run uses a 4-hour timeout with 32 CPU cores and 48GB RAM, and results are averaged over 5 runs. We adjust the Qwen chat template to ensure consistency between training and inference (https://huggingface.co/ornith-ai/Ornith-1.0-397B/blob/main/chat_template.jinja) and modify Harbor to align with vLLM’s reasoning_content key.
• Terminal-Bench 2.1 (Claude Code): We evaluate Terminal-Bench 2.1 using Claude Code 2.1.126 with parser=json, temperature=1.0, top_p=1.0, max_new_tokens=131072. Results are averaged over 5 runs. Again, the Qwen chat template needs to be modified.
• SWE-Bench Verified, Pro and Multilingual: using OpenHands harness with temp=1.0, top_p=0.95, and 256k context window.
• SWE Atlas QnA, RF, TW: using mini SWE agent harness with temp=1.0, top_p=0.95, 128K context window. Results are averaged over 5 runs.
• NL2Repo: with temperature=1.0, top_p=1.0, 400K context, 48K output and anti-hacking filters.
• ClawEval: An agentic code benchmark over real-user task distributions; temp=0.6 and 256K context.

The familiar image of spacetime as a cone-shaped map of past and future is often attributed to Albert Einstein. But its mathematical and visual foundation owes much to Hermann Minkowski, Einstein’s former professor.

Einstein’s 1905 special relativity changed how physics understood motion, light, and time. Minkowski then recognized that the theory becomes far clearer when space and time are treated as one unified four-dimensional structure: spacetime. In a famous 1908 lecture, he argued that space and time separately were “doomed to fade away into mere shadows,” while only their union retained an independent reality.

Minkowski died in 1909, before Einstein completed general relativity. Yet his framework became indispensable to the theory Einstein finalized in 1915 and published in 1916.

What the Light Cone Shows

A light cone is a diagram of causal possibility around a particular event—such as this moment, here and now.

• The past light cone contains events that could have sent light or other causal influences to you.
• The future light cone contains events you can in principle influence, provided no signal travels faster than light.
• Events outside either cone are “spacelike separated”: there is no time for a light-speed-or-slower signal to travel between them.

In the diagram, time is usually drawn vertically and space horizontally. The cone’s surface represents light itself. Your own path through spacetime—your “worldline”—stays inside the cone, because massive objects travel slower than light.

Minkowski used spacetime diagrams and described the conical separation between past, future, and causally disconnected regions. The light-cone structure remains fundamental in both special and general relativity.

What Einstein Added

Minkowski spacetime describes special relativity: a universe without gravity, or a sufficiently small region where gravity can be neglected.

Einstein’s general relativity made a deeper move. It treats gravity not as a conventional force acting across space, but as the curvature of spacetime caused by energy and matter. Planets, light, clocks, and freely falling objects follow paths shaped by that geometry.

So it is fair to say:

• Einstein introduced special relativity in 1905.
• Minkowski supplied the four-dimensional spacetime formulation that made special relativity geometrically powerful.
• Einstein developed general relativity, where spacetime itself can curve and evolve in response to matter and energy.

Minkowski’s framework is therefore not a side note to Einstein’s work; it is part of the language through which modern physics expresses relativity.

Does Relativity Prove the Future Has Already Happened?

This is where physics and philosophy meet.

Relativity does undermine the idea of one universal cosmic “now.” Two observers moving relative to each other can disagree about which distant events are simultaneous. There is no observer-independent plane slicing the whole universe into an objectively shared present.

That feature motivates the block universe, also called eternalism: the view that past, present, and future events all belong to a four-dimensional spacetime whole. In this picture, “now” is not a universal boundary sweeping across reality; it is a local feature of an observer’s experience and position in spacetime.

But an important distinction matters: the light cone does not itself prove that “the future already happened.” A light cone primarily maps causal structure—what can affect an event and what it can affect. The block universe is a philosophical interpretation strongly motivated by relativity’s lack of absolute simultaneity, not a single experimental conclusion forced by the diagram.

A More Accurate Way to Say It

A more precise formulation would be to say, not “the past, present, and future already happened,” but rather:

Relativity describes the universe as a four-dimensional spacetime in which there is no single, observer-independent present. This has led many physicists and philosophers to interpret time as a “block,” where past, present, and future are equally part of spacetime — but that interpretation remains philosophically debated.

Short answer: A token is the smallest chunk of text an AI model processes—roughly ¾ of a word in English. Every prompt you type and every reply you get back is billed, limited, and generated in tokens, not words or characters. Understanding this one concept explains why AI costs money, why long chats slow down, and why answers sometimes get worse the longer you talk.

Most explanations of AI tokens jump straight to “optimization "tips"—shorter prompts, fewer uploads, cheaper models. But almost none of them explain what a token is first. That’s backwards. You can’t optimize a system you don’t understand, so let’s start with the basics.

What Is a Token, Exactly?

A token is a fragment of text—sometimes a whole word, sometimes just a piece of one. Models don’t read letters or words the way people do; they read these fragments, called tokens, and convert each one into a number before any “thinking” happens.

Here are a few specific guidelines:

• Short, common words (“the,” “is,” “cat”) are usually one token.
• Longer or unusual words get split into pieces. “Understanding” might become “Under” + “stand” + ""ing"—three tokens for one word.
• Punctuation, spaces, and symbols also often count as their own tokens.
• As a rough English-language estimate: 100 tokens ≈ 75 words.

So when someone says a model has a “128,000-token context window,” they’re really describing how much text—roughly 95,000 words, or a 300-page book—it can hold in memory at once.

How Does Tokenization Actually Work?

Every AI conversation follows the same round trip:

Your text → split into tokens → converted into numbers → processed by the model → new numbers generated → converted back into the words you read.

That splitting step is called tokenization, and it isn’t standardized across companies. GPT, Claude, and Gemini each use their own tokenizer, which is why the exact same sentence can quietly cost a different number of tokens depending on which model you’re talking to.

The Different Types of Tokens You’re Actually Billed For

Not every token costs the same, and not every token is even visible to you. Here’s what shows up on an API bill or eats into a usage limit:

Input tokens—everything you send: your message, the full conversation history, and any attached files.

Output tokens—what the model writes back. Because generating text takes more computation than reading it, output tokens typically cost three to five times more than input tokens. This is the single most overlooked cost driver in AI usage.

Cached tokens—many providers now support prompt caching, which stores a repeated block of context (a system prompt, a long document, or a style guide) so it doesn’t have to be reprocessed on every message. This can cut costs by 50–90% for repetitive workflows.

Reasoning tokens—newer “thinking” models generate an internal chain of reasoning before producing a visible answer. You’re often charged for these hidden tokens even though you never see them, which is why a short answer can sometimes carry a surprisingly high bill.

Multimodal tokens—images, audio, and video all get converted into tokens before a model can process them. A single screenshot can consume more tokens than several pages of plain text.

System and structural tokens—background instructions, tool calls, and formatting markers that keep the model’s internal structure intact. You rarely see them, but they still count against your limits.

Why Do Long Conversations Get Pricier Over Time?

This is because most chat interfaces have no persistent memory between messages. To stay coherent, the model rereads the entire conversation history on every single turn—not just your newest message.

That means cost compounds as a chat grows:

TurnApproximate tokens processedMessage 1~50 (just your prompt) Message 2~150 (prior exchange + new message) Message 5~400+ (full history so far) Message 15~2,000+ (long history dragged forward every time)

Your new message might be the exact same length each time—but the bill keeps climbing because everything behind it keeps growing.

What Actually Drains Your Token Budget Fastest

• Long, sprawling conversations—full history resent every turn
• Large documents or PDFs—a single page can run 800–1,000 tokens, so a 30-page file can burn 30,000 tokens before you’ve asked a question
• Image uploads—one screenshot can outweigh several pages of text
• Vague prompts—the model needs more back-and-forth context to get it right
• Agentic or multi-step workflows—every tool call and reasoning step adds tokens
• Long AI responses—output is the expensive side of the ledger

Why Does the AI Get Worse the Longer I Talk to It?

This is the pattern almost everyone has noticed without knowing why:

• “It gave great answers, then got generic.” The context window filled up, leaving less room for the model to reason clearly.
• “I hit my usage limit way earlier than expected.” File or image uploads silently consumed thousands of tokens.
• “The conversation got noticeably slower.” Every message was dragging a growing history behind it.
• “A new chat suddenly gave a sharper answer.” Zero history means zero token baggage—the model has its full context window available again.

How to Reduce Token Usage: 8 Practical Fixes

1. Start a new chat when you switch topics. Unrelated history is dead weight that gets billed on every future turn.
2. Ask for shorter answers. Since output costs more than input, “answer in three sentences” saves more than trimming your prompt ever will.
3. Paste only the relevant section of a document, not the whole file, when you need help with just one part.
4. Skip image uploads unless the model truly needs to see something—describing it in words is almost always cheaper.
5. Use context-compacting tools (like /compact in Claude Code) to summarize history instead of starting over or letting it balloon.
6. Turn on prompt caching if you repeatedly send the same system prompt, reference document, or instruction set.
7. Match the model to the task. Route simple formatting, translation, or one-line questions to a lighter, cheaper model, and save the most capable model for complex reasoning.
8. Request structured output (JSON, tables, bullet lists) when building with an API—it's typically more token-efficient than free-form prose.

Frequently Asked Questions About AI Tokens

Is a token the same as a word? No. A token is roughly ¾ of a word on average in English. Short words are usually one token; longer or unusual words split into multiple tokens.

Why do output tokens cost more than input tokens? Generating new text requires more computation per token than reading existing text, so most providers price output at 3–5x the input rate.

Does starting a new chat actually save tokens? Yes. Because chat interfaces resend the full conversation history with every message, a new chat resets that cost to zero and frees up the model’s entire context window.

What is a context window? It’s the maximum number of tokens (input plus output combined) a model can hold and process in a single conversation at once. Once you exceed it, the oldest content typically gets dropped or summarized.

Do images and PDFs really use that many tokens? Yes—a single page of a PDF can run 800–1,000 tokens, and one image can consume more tokens than several pages of plain text, since the file has to be converted into token form before the model can “see” it.

The One Thing to Remember

Tokens are the invisible currency behind every AI interaction—they quietly determine your cost, your speed, your usage limits, and why long conversations start to feel less sharp. You don’t need to count every one. You just need the instinct that every word sent and every word returned is being metered, and the meter runs faster the longer the conversation goes.

Once you think in tokens, AI stops feeling unpredictable—the limits, the costs, and the slowdowns all make sense.

Hope this helps,
Dan

Google has just announced a major update to Gemini that will change the way you work, create and get things done. If you’ve been waiting for AI to actually do stuff, not just talk about doing stuff, now’s the time.

Quick Look at the Latest Gemini Updates
Google had five major updates that affected various aspects of how you use AI. New user interface. Video creation tools. 24/7 AI agent Spark. Upgraded Google Flow features. New model called Gemini 3.5 Flash.

The most recent updates were announced at Google I/O 2026 in May, with more features being released through June and July 2026.

Why These Updates Matter in Your Daily Work
Chances are you spend hours building presentations, editing videos, or managing your inbox. These tools were built by Google to cut that time down. Slides now lets you create an entire presentation in minutes, not hours, with its new deck generation capability.

In Docs, Gemini now manages comment threads and generates visuals without ever leaving the document. That means you don’t have to bounce back and forth between different tools to do your work.

Google I/O 2026: The 5 Big Announcements
Google released five updates that really make sense for actual work, and everyone is surprised. Here’s what they each do and why it matters.

New Gemini User Interface and Organization of Messages
The Gemini app has a new look, which it’s calling Neural Expressive. Smoother animations, haptic feedback, and better colors. The chat organization is better with separate spaces for images, videos, and your library of created content.

Gems also got a facelift. Better organize your custom AI assistants and find what you’ve created in the past. The entire interface feels less cluttered and more like how you actually think about your work.

Make Videos with Gemini Omni
Here is where it gets crazy. You can create videos with text prompts. Want an 8-second video of a professor explaining E=MC squared? You type it in, select landscape or portrait, and Gemini makes it happen.

The videos have 2 or more angles, sound and sometimes even talking. You can then edit them like "change him to wearing a navy blue suit," and it actually does that while keeping everything else the same.

You want to upload your own images or avatars to put yourself in scenes. The tool bundles images, text and videos into a single workflow. It’s a game changer for content creators who need B-roll or quick video assets.

Gemini Spark: Your AI Agent, 24/7
Your phone and laptop can be off; Spark still works. It runs in the background, doing things you ask it to do. You can set schedules, create skills such as ghost writing or organizing files in Google Drive, and automatically track incoming leads.

The agent lives in your inbox and any other connectors you set up. It can surf the web, make reservations,, and buy items on many sites. It’s a comparison tool that helps you make bookings without having to hover over each step.

Spark isn’t supposed to start until the week after the announcement, but it’s built for Google AI Ultra subscribers. Available first in Gmail and Chat.

Google Flow Character Creation Enhancement
Google Flow is a creative companion to help you brainstorm, prompt and edit. Create and share your own custom tools in Flow. Gemini Omni Flash can be used on the go with the mobile app.

The big deal here is creating the character. You build your character and describe it, or make it look like yourself. This creates AI avatars inside the app. You direct scenes, work with image editors and add post-processing effects like low-fi and glitch.

It’s essentially an AI version of video creators and editors all in one app. Add text, effects,, and resize videos with simple commands.

AI Studio & Gemini 3.5 Flash
Gemini 3.5 Flash is the most capable model for sustained frontier performance on agentic encoding tasks. It is 4x faster than other leading models, measured in tokens of output per second.

Run autonomous agents on a remote Linux environment hosted by Google using the API. AI Studio now supports this. Create customer support agents, repo maintainers, document processors, and data analysts bywriting what you want the agent to do.

This is free, which is pretty crazy for what you get.

Gemini 3.5 Flash Features Breakdown
It’s the default model in the Gemini app and in AI Mode in Google Search. It is better at long running tasks. It is also better at token efficiency.

Speed & Performance Improvements
The model processes answers 4 times faster than any other top model. That speed matters when you’re running agents or doing sustained work that takes time.

Google pushed Gemini 3.5 Pro from June to July for longer-running tasks and AI agents. They are listening to more feedback from testers to get it right.

Skills for Using the Computer
Gemini 3.5 Flash combined with computer use. You build custom agents that observe, think,, and do in desktop, mobile,, and browser contexts.

This improves performance for enterprise automation tasks and long-horizon tasks. Picture software testing that never ends andknowledge work that can hum along without you hovering over its shoulder.

Tasks for enterprise automation
If your business processes take hours or days, this update is for you. Agents can manage multi-step processes that would typically require a human to check in at each step.

Gemini Omni Flash: Making Video Creating Easy
Gemini Omni Flash is a native multimodal model for enterprises and developers. For the first time ever, it allows you to develop custom, dynamic video workflows.

How to Generate Videos from Text Prompts
Type what you want to see. The model produces video by combining images, text,, and motion. You can add aspect ratio, length, and style to your prompt.

This works in Google Vids, Google Flow,, and the main Gemini app. Different tools give you different levels of control on the output.

Edit Existing Videos with AI
Edit phone video or AI clips with text prompts. You can test, repurpose, or extend video assets without traditional editing tools.

The tool understands context, and makes changes that are of the style of the original text. That’s like having an editor who understands when you say, ‘make it punchier’ or ‘slow down the middle section.’

Avatar & Character Integration
Generate new videos with your avatar that looks and sounds like you. You get presenter effect, no studio time or camera setup

In Google Vids, your avatar becomes a digital presenter that storyboards slides or writes scripts. Re-script your avatar anytime or generate AI voiceovers in 8 languages including Japanese, Korean, Spanish and French.

Google Workspace Gets Big AI Boost
In July 2026, the Workspace Drop brought AI features to Slides, Vids, and Docs, changing the way teams work on content together.

Gemini for Presentations in Google Slides
The deck creation feature will create complete decks in minutes. You tell it what you want, and it looks at existing decks to match the formatting of your brand.

Gemini pulls context from your Google Docs, Sheets, PDFs, or past decks. It asks clarifying questions and builds a complete plan. The result is a fully editable deck that meets your needs.

Gemini in Google Videos for Video Editing
Digital presenters can automatically storyboard your slides or generate scripts. Your avatar is a speaker that you can re-script at any time.

Vids’ Gemini Omni Flash lets you edit any video footage with simple text prompts. Build new videos with avatars that look and sound like you.

Google Docs Comment Action Options
Gemini in Docs summarizes comment threads and generates responses to feedback. It suggests document changes that you can click to accept.

It can also generate and edit images, diagrams, and infographics using natural language prompts. It knows your document context. It generates relevant visuals. Edit multiple visuals simultaneously without ever leaving Docs.

Android 17 & Pixel Drop Features:
With Android 17, you’ll find features that will make it easier to multitask and create content on your phone.

Multitasking and Floating App Windows
Run multiple apps at once with floating app windows. Screen Reactions lets you record picture-in-picture for tutorials or quick how-to videos.

Foldable gaming has an optimized layout that better utilizes the larger screen. These updates will come first to Pixel devices, then to other eligible Android devices through 2026.

AI-Powered Video & Music Creation
The June Pixel Drop brings AI-powered video and music creation tools. You can create content directly from your phone without needing desktop software.

Voice translation expanded to assist in working with folks in other languages. The practical features you actually use include custom voicemail greetings and automated emergency notifications.

Improved security and biometrics
Enhanced security features include the ability to lock a lost phone with your biometrics. This adds an additional layer of protection if your device gets lost or stolen.

Google Home Speaker, New for Gemini
Google announced a new smart speaker designed specifically for Gemini. It makes talking to your home assistant more natural.

Conversations in Nature Without Rigid Rules
You don’t have to use strict commands anymore. The speaker feels like a real person and is better at dealing with context than earlier voice assistants.

It can answer complex questions and handle multiple requests simultaneously. The speaker remembers what you were just talking about, so you don’t have to repeat yourself all the time.

Multiple requests and memory functions
The speaker can reply to more than one request in a single dialogue. You can ask follow-up questions about previous parts of your chat.

The memory feature makes the speaker less of a tool and more of a person that actually listens to what you’re saying.

Gemini 3.5 Live Translate: Overcoming Language Barriers
The new audio model of live speech-to-speech translation can automatically detect more than 70 languages. It retains the speaker’s natural inflections and eliminates awkward pauses.

70+ Languages Supported
The model covers the most important business and travel languages. Speak fluently on the fly in multicultural calls, meetings, or trips.

It removes language barriers in seconds, instead of waiting for translation apps to catch up.

Preservation of Natural Intonation
The translation retains the tone and rhythm of the speaker’s voice. It doesn’t feel robotic like previous translation tools.

The feature is rolling out in the Gemini Live API, Google AI Studio, and the Google Translate app.

Subscription Tiers and Price Changes
Google has revealed new pricing tiers for its AI Ultra, making it more affordable. I/O 2026

Google AI Ultra price drops
The entry price for Google AI Ultra is $200 per month. It’s now cheaper than its previous price of $250 a month.

Get the latest Gemini features, including AI Inbox, for less. This makes the advanced features available to more people.”

New $99 Ultra Option
Google also added a cheaper Ultra option for $99 a month. This tier gives access to key features without the full Ultra price.

The change in price is a sign that Google wants to get more people to use Gemini in their day-to-day work.

What’s next for Google?
Google has more updates lined up for the rest of 2026. Here’s what to look for.

Gemini 3.5 Pro pushed back to July
Gemini 3.5 Pro was delayed from June to July 2026. Google is working to improve the model with more tester feedback before it’s released.

The focus is on performance improvements for long-running jobs, AI agents, and token efficiency. This delay should lead to a better final product.

Chrome Integration & WebMCP;
Chrome 149 will launch an experimental origin trial for support for WebMCP APIs in Gemini.

That means AI features built right into your browser experience. Browse and work with AI without leaving the app.

How to Get These Features Now
Most of these features are now rolling out depending on your location and subscription.

Scaling to Additional Areas
Gemini in Chrome is rolling out to users across the UK and other regions.

Some features are available globally today, and others will roll out gradually through 2026.

Availability of Workspace Customers
We’re now rolling out the July 2026 Workspace updates to Google Workspace customers.

Docs’ Gemini now supports 11 new languages, including Mandarin, Dutch, Malay, Hebrew, Polish, Turkish, Czech, Indonesian, Swedish, Danish, and Norwegian.

Data regions are now supported in the Gemini app, enabling admins to select whether AI data is processed and stored in the EU or US geographically.

FAQs
When will Gemini Spark be released?
The Gemini Spark doesn’t even launch until the week after the announcement. It will be available to Gmail and Chat users in the U.S. who subscribe to Google AI Ultra.

Gemini 3.5 is Free or Paid?
The Gemini 3.5 Flash is the default model for the Gemini app and in Google Search, AI Mode.

Some advanced features require a Google AI Ultra membership. The Ultra plan costs $99 a month and gives you access to key features, while the $200-a-month tier unlocks everything.

Is Gemini compatible with my phone?
Yes. The Gemini app is available on Android and iOS. The new Neural Expressive design was rolled out globally across Gemini’s desktop, Android, and iOS apps.

Google Flow has a mobile app to create content on the go.

Privacy and Data Security?
Google’s Gemini app now includes support for data regions, offering admins fine-grained control over where their AI data is processed and stored.

The firm filed a civil lawsuit against cybercrime operations that employ AI in scams. They’re pushing for bipartisan legislation to fight AI-powered scams and using AI tools to combat them.

Closing Thoughts on Google’s AI Path
Google’s updates prove they’re serious about building real, functional AI—not just a chatbot. 24/7 agent. Video creation. Workspace integration. Signs of AI doing the things you don’t want to do.

The price changes make the more advanced features more accessible. If you are in one of the regions where these features are rolling out, now is the time to try it out in your real workflow.

Open up a monthly bank statement, and the story is usually the same. A tiny recurring charge here, a subscription there, and suddenly a team is paying hundreds of dollars a month for tools people barely open. It functions like a permanent tax on just getting work done. From scheduling meetings to backing up photos, signing documents, or drawing diagrams, digital operations have become entirely rental-dependent.

But a shift is happening in the software world. Almost every major paid platform now has an open source equivalent that does the exact same job, built transparently on GitHub for anyone to deploy. These are not buggy weekend side projects. They are massive repositories backed by hundreds of thousands of developers and run in production by real corporations. By swapping out a commercial logo for a self-hosted alternative, companies and individuals are dropping their software bills entirely to zero.

The transition requires a change in mindset. Instead of renting software forever, users set it up once on their own server and own it completely. Here is a look at the top ten open source repositories currently replacing expensive monthly subscriptions across the tech industry.

10. Documenso

Sending contracts for digital signatures typically leads straight to DocuSign, which gates basic features behind a perpetual per-user monthly fee. Documenso offers a complete open-source alternative. Users can upload a PDF, place signature fields, and generate a secure signing link for clients. The platform provides real digital signatures and a full, tamper-evident audit trail logged directly in the browser, completely bypassing the seat-based tax.

9. Cal.com

Calendly built a massive business around a singular, highly useful feature: picking a time to book a meeting. Cal.com takes that exact functionality and allows users to host it on their own domain without an upsell wall. Because the underlying code is fully open, businesses can customize the booking flow to match their exact structural needs. It handles time zones, calendar syncs, round-robin team distributions, and embeds effortlessly anywhere, unlocking features that paid tools reserve for premium tiers.

8. PostHog

Product analytics tools like Amplitude and Mixpanel often hook startups with generous free tiers, only to hit them with massive, scaling invoices the moment user traffic grows. PostHog prevents that event-based metering panic. It packages product analytics, session replays, feature flags, user surveys, and A/B testing into a single repository. Teams can track funnels, retention curves, and watch actual session recordings of users clicking through their apps without worrying about scaling data costs.

7. Penpot

Figma remains an industry standard for design, but corporate ownership changes have left many teams uneasy about keeping their entire design systems locked in a proprietary ecosystem. Penpot serves as a web-based, open source alternative that cannot be bought out from under its users. It bridges the gap between design and development by generating clean, production-ready code straight from the canvas, removing the need for clunky export plugins or middleman paywalls.

6. NocoDB

Airtable made relational databases look as beautiful and friendly as spreadsheets, but the costs scale rapidly at twenty dollars or more per user. NocoDB solves this by connecting directly to an existing database, such as Postgres or MySQL, and turning it into an interactive grid. Users get access to kanban boards, calendars, relations, and public forms while keeping full custody of their data rather than storing it in a third-party cloud.

5. Coolify

Platforms like Vercel, Heroku, and Netlify provide an incredibly smooth push-to-deploy experience, but they are notorious for sudden bandwidth bills and steep team pricing. Coolify recreates that exact magical deployment flow on an individual’s own server. Users connect a Git repository, and Coolify handles the building, shipping, automated backups, and SSL certificates. The cost remains entirely flat and predictable, tied only to the server it runs on.

4. Excalidraw

When teams need to map out an architecture sketch or brainstorm on a digital whiteboard, apps like Miro and Lucidchart demand monthly editor subscriptions. Excalidraw brings that casual, hand-drawn whiteboard aesthetic straight to the browser for free. It supports instantaneous live collaboration, simple shape drawing, and quick one-click exports, making it a go-to tool for developers trying to explain complex ideas clearly.

3. Supabase

For years, Google Firebase was the default choice for quick backend development, but it locked mobile and web apps into a proprietary database ecosystem. Supabase entered the market as an open source answer, offering a dedicated Postgres database alongside authentication, file storage, and instantly generated APIs. Developers get the speed of a ready-made backend without the fear of vendor lock-in, meaning they can export their entire setup and walk away whenever they want.

2. n8n

Visual automation tools like Zapier charge users strictly by the task, meaning a successful, high-volume automation workflow can quietly cost hundreds of dollars a month. The open source tool n8n operates as a self-hosted node builder where running ten thousand workflows costs the exact same as running ten. Users drag triggers, actions, and even custom AI steps onto a visual canvas, running the entire operational sequence safely on their own hardware.

1. Immich

Immich has become one of the most celebrated projects in the self-hosting landscape because it replaces Google Photos with stunning accuracy. The mobile application automatically backs up camera rolls, while the web dashboard offers facial recognition, instant semantic search, and an identical scrolling timeline. The critical difference is that the photos remain entirely on private hardware, protected from automated corporate scanning.

The Real Cost of Free Software

While these tools eliminate monthly credit card charges, a basic truth remains: free software simply shifts the currency of payment from money to time. When self-hosting, someone has to manage the server, install version updates, configure secure backups, and fix the system if it goes offline unexpectedly. A commercial subscription essentially buys the luxury of never thinking about infrastructure maintenance. Furthermore, tools like n8n operate under fair-code licenses, meaning they are entirely free for personal use but require payment if repackaged and resold as a commercial service.

For teams with no operations personnel or zero interest in server upkeep, paying the premium for hosted versions is entirely reasonable. In fact, many of these open-source projects offer official cloud versions, allowing users to fund the original creators directly rather than paying a software middleman. The broader takeaway is not that commercial platforms are inherently bad, but rather that developers and businesses finally have a legitimate choice. Replacing just a few key pieces of the modern software stack can easily save thousands of dollars annually, proving that some of the most powerful tools in tech are no longer for sale. They are simply sitting in a public repository, waiting to be cloned.

For more breakdowns on open source systems and deployment strategy, check out the full video on the Cloud Codes YouTube channel.

This isn’t really much of an article, but it does have a TON of VALUE!!!!

Meme Generator

Meme Generator
I built an easy-to-use meme generator that adds a little motion for a better user experience. Just replace the default image with any photo or AI-generated image of your choice and start playing around. You can also use it for product images. ENJOY:

Motion Studio - Fashion Short

Why Memes Are the Most Underrated Growth Tool in Your Content Strategy

If you’re building an audience in 2026 and memes aren’t part of your content mix, you’re leaving one of the cheapest, fastest, most human tools on the table.

Everyone knows memes are funny. Fewer people treat them as strategy. That’s the gap this article is here to close.

Memes Are a Trust Shortcut

Building an audience is really about building trust at scale. The problem is trust usually takes time — people need to see you show up, get a feel for your voice, decide you’re not selling them something shady.

Memes compress that timeline. A meme that actually lands tells a stranger, in under two seconds, “this account gets it.” It signals shared context, shared humor, shared frustrations. That’s a relationship shortcut no polished explainer video can replicate.

They’re Built for the Way Algorithms Actually Work

Every platform you listed — YouTube, TikTok, Instagram, Facebook, Threads, Telegram, X, Reddit — rewards the same underlying signals: quick engagement, high completion rate, shares, and saves. Memes are practically engineered for all four.

• Quick engagement: a meme is understood instantly, so the like/comment/share decision happens fast.
• High completion: there’s nothing to “finish” — the whole thing is consumed in one glance, which platforms read as a strong retention signal.
• Shareability: memes are made to be sent to a specific friend with “this is so you.” That one-to-one share is gold for distribution because it moves your content into DMs and group chats, not just feeds.
• Saves: a genuinely funny or relatable meme gets bookmarked to use later, which many algorithms weight heavily.

You don’t get all four of those consistently from long-form content. You often get them reliably from a good meme.

Memes Lower the Barrier to Your First Follow

Nobody subscribes to a newsletter from a stranger. Plenty of people will follow an account that made them laugh twice in one scroll session.

Memes work as a low-commitment audition. The person isn’t agreeing to your worldview or your 20-minute video essay — they’re just enjoying a ten-second dopamine hit. But every one of those hits is a rep. Enough reps and the algorithm starts showing you to that person more often, and that’s when familiarity turns into an actual follow.

They Reveal What Your Audience Actually Cares About

Post enough memes and you get something better than a survey: real behavioral data. Which jokes get shared versus just liked? Which format — the “expanding brain,” the two-panel comparison, the reaction image — gets the most comments? Which topics get people tagging their friends?

That data quietly tells you what your audience is actually starved for, which you can then feed back into your bigger content — the videos, the posts, the products, the newsletter. Memes become a free, constant focus group.

Consistency Without Burnout

One of the biggest reasons creators quit is production fatigue. Long-form content takes hours. Memes take minutes. That means you can post daily, sometimes multiple times a day, without torching your energy — and daily presence is exactly what most of these algorithms want to see before they start trusting your account with real reach.

A meme account with unglamorous consistency will often outgrow a polished account that posts once a month. Platforms reward the pattern of showing up, not just the occasional home run.

Memes Travel Across Platforms Natively

A meme made for Instagram works on Threads. A meme made for X works on Facebook. A meme made for Reddit works on Telegram. Unlike a 60-second TikTok that needs re-editing for YouTube Shorts, or a long-form video that needs a totally different cut for X, memes are format-agnostic. One asset, ten platforms, ten chances to be discovered.

This is a massive efficiency win if you’re trying to build presence everywhere at once instead of picking a single platform and hoping.

The Caveat: Memes Build Reach, Not Loyalty — On Their Own

Here’s the honest part. Memes are extraordinary at the top of the funnel — reach, discovery, first impressions. They are not, by themselves, great at building deep loyalty or monetizable trust. Nobody buys a course from a meme account with no other substance behind it.

The creators who actually turn meme reach into a real audience use memes as the front door, then walk people toward something with more of their voice and expertise — a newsletter, a podcast, longer videos, a product. The meme gets the stranger to stop scrolling. What happens next is what turns them into a fan.

The Takeaway

Memes aren’t a distraction from “serious” content strategy — they are content strategy, just tuned for the reality of how modern platforms distribute attention. They build trust fast, they feed the algorithm what it wants, they lower the cost of a first follow, and they hand you free audience research in the process.

If you’re trying to grow anywhere from TikTok to Reddit to Telegram, don’t treat memes as filler between “real” posts. Treat them as the on-ramp. Then build the road that keeps people driving.

There is a noticeable shift happening right now in the world of internal energy arts. For a long time, the trend leaned toward chasing complex, advanced forms. Everyone wanted the flashier, more exotic techniques. But recently, a collective consensus across online forums and practitioner communities shows that people are stepping back. They are realizing that without foundational somatic mastery, the advanced stuff simply does not work. The focus is moving toward a highly structured sequence that strips away the noise and prioritizes two core elements: the Inner Smile and foundational breathing exercises.

Sequence and Structure

The community is waking up to the fact that order matters. Skipping the basics to get to the advanced levels is proving to be a dead end for most. The current movement emphasizes a clean, logical progression. It is about building a stable physical and energetic foundation before attempting to channel larger amounts of energy. Practitioners are finding that mastering a simple, structured routine yields far better results than collected sequences of half-learned, complex movements.

The Inner Smile Layer

This brings us to what many are calling the Inner Smile layer. A common mistake among practitioners is trying to force Qi or build internal heat right out of the gate. The feedback from modern forums shows that this aggressive approach usually leads to stagnation, blocks, or plain old frustration. The human body requires a neurogenic state of safety before it allows energy to move smoothly. By practicing the Inner Smile first, a practitioner signals to their nervous system that it is safe to drop its guard. Without that initial layer of internal peace, trying to force energy is like hitting the gas pedal while the emergency brake is fully engaged.

Breath as the Regulator

Once you establish that state of safety, breath becomes the regulator. Modern practitioners heavily analyze deep, rhythmic abdominal breathing as the literal engine of the entire practice. It acts as the physical bridge between the conscious mind and the autonomic nervous system. A person cannot easily think their way into a lower heart rate or a calmer state, but they can breathe their way there. This specific type of deep abdominal breathing moves the practice from a vague mental exercise into a true somatic reality, anchoring the energy and giving the body a reliable mechanism to regulate itself.

For those looking to dive deeper into how these foundational layers fit together, resources are available to help map out the daily practice.

• You can find a complete breakdown of these core principles and training philosophies at developyourenergy.net.
• To get started on the exact sequencing mentioned here, read the detailed guide on the inner smile and six healing sounds practices.
• If you prefer a visual and auditory walkthrough to follow along with, check out this instructional guide on YouTube.

The landscape of SEO in 2026 is fundamentally different from what it was just a few years ago. If you talk to anyone running a website right now, they’ll tell you the same thing: traffic is down, and the old playbook is broken. When a Google AI Overview pops up at the top of a search page, the number one organic result loses up to 58% of its clicks.

People aren’t just searching differently; search engines are processing information differently. But instead of throwing your hands up in defeat, the trick is learning how to use these shifts — and tools like NotebookLM — to build a modern AI SEO strategy.

The New Reality of How People Find Things

Think about the last time you bought a product. You probably started on Google, clicked a couple of lists, ignored half of them, and then went straight to Reddit to see what actual humans were saying. Maybe you watched a YouTube review or checked Amazon reviews while standing in the store aisle.

AI assistants like ChatGPT, Perplexity, and Google’s AI Overviews learn exactly the same way we do. They crawl forums, niche blogs, YouTube transcripts, and review sites to decide who to recommend.

Because of this, modern AI SEO requires focusing on three specific shifts:

1. The Power of “Query Fanout” and Branded Mentions

When someone types a broad query into an AI assistant, the AI doesn’t just pull an answer from memory. It breaks that single prompt down into dozens of smaller subqueries. If someone asks for the best smartphone, the AI secretly searches for specific comparisons, low-light photo tests, and zoom capability reviews, then stitches the answers together.

If your brand is mentioned across the pages that rank for those smaller subqueries, you get included in the final AI response. Branded mentions on third-party sites now have a stronger correlation with AI visibility than traditional metrics like domain rating or backlinks.

2. Targeting Action Over Information

AI is incredibly good at answering simple questions. It can explain compound interest or calculate a BMI instantly, meaning informational blog posts are losing organic traffic fast.

Where traditional SEO still completely dominates is action-driven queries. If someone searches for a “mortgage calculator” or a “snow removal service,” an AI summary can’t satisfy them. The user needs to click a link, use a tool, or make a transaction. Focusing your content on transactional intents and interactive tools keeps your traffic safe.

3. Overcoming the Indexing Bottleneck

With AI making it incredibly easy to generate thousands of pages, Google has gotten incredibly picky about what it actually crawls and indexes. If your content isn’t indexed, it doesn’t exist. Modern workflows rely on automated indexing tools and APIs plugged directly into content pipelines to ensure new pages hit the search results immediately.

Enter NotebookLM: Your Strategic SEO Command Center

With the rules of the game clearly defined, how do you actually execute this without drowning in research? This is where NotebookLM becomes your secret weapon.

Most people use AI as a simple text generator, which results in generic, easily filtered content. NotebookLM is different because it acts as a closed-loop synthesis tool grounded entirely in the sources you give it.

Here is how to use it to drive a modern AI SEO strategy:

Reverse-Engineering the Competition

Instead of guessing what kind of content will get you cited by AI engines, you can use SEO tools to export reports of the exact domains and pages currently being cited in your niche. Drop those reports, along with competitor top-performing pages and forum transcripts, directly into a NotebookLM notebook.

Ask the notebook: “What specific subqueries are AI assistants pulling from these sources to answer broad questions in our niche?” This reveals the exact “query fanout” map you need to target.

Identifying the “Action” Gaps

Upload your current keyword list and search search data into NotebookLM. You can prompt it to sort through the noise: “Filter these topics based on whether an AI overview can fully satisfy the user’s intent. Highlight the queries that require a functional tool, specific local action, or a transaction.” This instantly filters out dead-end informational keywords and gives you a high-value content roadmap.

Creating Hyper-Grounded Content Briefs

AI SEO fails when it sounds like a robot summarizing a textbook. To stand out to both human readers and search algorithms, your content needs proprietary data, unique case studies, and real depth.

By loading NotebookLM with your internal product data, original customer case studies, and unique brand viewpoints, you can generate comprehensive content briefs. The AI won’t pull generic fluff from the public web; it will build structures tightly aligned with your actual brand expertise, ensuring your content has the high-quality substance that search engines want to index.

The Bottom Line

The old version of SEO isn’t entirely dead, but it has evolved into a broader game of digital omnipresence. Success today means building a brand that gets talked about on review sites, forums, and niche blogs, while protecting your own site with action-oriented content. By using NotebookLM to analyze the landscape and ground your strategy in real data, you can stop guessing what the algorithms want and start building an authoritative presence that commands attention.