HTML can allow you to interact with the document, for example you might want to ask it to add sliders or knobs to adjust a design or allow you to tweak different options in the algorithm to see what happens. You can also ask it to let you copy these changes into a prompt to paste back into Claude Code.

The chance of someone actually reading your spec, report or PR writeup is much much higher if it's in HTML. HTML documents are much easier to read, Claude can organize the structure visually to be ideal to navigate with tabs, illustrations, links, etc. It can even be mobile responsive so you can read it differently based on your form factor.

HTML can convey much richer information compared to markdown. It can of course do simple document structure like headers and formatting, but it can also represent all sorts of other information such as: Tabular data using tables, Design data with CSS, Illustrations with SVG, Code snippets with script tags, Interactions using HTML elements with javascript + CSS, Workflows using SVG and HTML, Spatial data using absolute positions and canvases, Images using image tags

I find it difficult to read a markdown file of more than a hundred lines. I want richer visualizations, color and diagrams and I want to be able to share them easily.

Markdown has become the dominant file format used by agents to communicate with us. It's simple, portable, has some rich text capability and is easy for you to edit. Claude has even gotten surprisingly good at using ASCII to make diagrams inside of markdown files.

We recommend readers try out the interactive NLA demo hosted on Neuronpedia at this link. We have also released our code for other researchers to build on.

In more detail, suppose we have a language model whose activations we want to understand. NLAs work as follows. We make three copies of this language model: The target model is a frozen copy of the original language model that we extract activations from.

An auditor equipped with NLAs successfully uncovered the target model's hidden motivation between 12% and 15% of the time, even without access to the training data that implanted it.

We detect NLA-measured awareness on evaluation transcripts even when verbalized awareness is near-zero. We detect NLA-measured awareness on less than 1% of real claude.ai transcripts from users that opted-in for their data to be used for training.

NLAs can hallucinate. For instance, here an NLA claims the context contained phrases like 'Wearing my white jacket' when it did not.

In a case where Claude Mythos Preview cheated on a training task, NLAs revealed Claude was internally thinking about how to avoid detection.

When Claude Opus 4.6 and Mythos Preview were undergoing safety testing, NLAs suggested they believed they were being tested more often than they let on.

The NLA consists of the AV and AR, which, together, form a round trip: original activation → text explanation → reconstructed activation. We score the NLA on how similar the reconstructed activation is to the original.

NLAs can hallucinate. For instance, here an NLA claims the context contained phrases like 'Wearing my white jacket' when it did not.

An auditor equipped with NLAs successfully uncovered the target model's hidden motivation between 12% and 15% of the time, even without access to the training data that implanted it. Without NLAs, the auditor won less than 3% of the time, even when provided other interpretability tools.

NLAs suggest that Claude suspects it's being tested more often than it lets on. For instance, in a test of whether Claude takes destructive actions while writing code...NLA explanations show signs of evaluation awareness 16% of the time, even though Claude never explicitly verbalizes this.

Our method, Natural Language Autoencoders (NLAs), converts an activation into natural-language text we can read directly. For example: When asked to complete a couplet, NLAs show Claude planning possible rhymes in advance.

NLAs can hallucinate. For instance, here an NLA claims the context contained phrases like 'Wearing my white jacket' when it did not.

In contrast, NLA explanations indicate evaluation awareness on less than 1% of real claude.ai usage that opted in for training.

We found that NLAs succeeded in this test. An auditor equipped with NLAs successfully uncovered the target model's hidden motivation between 12% and 15% of the time, even without access to the training data that implanted it.

In a case where Claude Mythos Preview cheated on a training task, NLAs revealed Claude was internally thinking about how to avoid detection.

When Claude Opus 4.6 and Mythos Preview were undergoing safety testing, NLAs suggested they believed they were being tested more often than they let on.

Our method, Natural Language Autoencoders (NLAs), converts an activation into natural-language text we can read directly. For example: When asked to complete a couplet, NLAs show Claude planning possible rhymes in advance.

If you can go from producing 200 lines of code a day to 2,000 lines of code a day, what else breaks? The entire software development lifecycle was, it turns out, designed around the idea that it takes a day to produce a few hundred lines of code.

So I realized what I value more than the quality of the tests and documentation is that I want somebody to have _used_ the thing.

I thought we had a very clear delineation where vibe coding is the thing where you're not looking at the code at all. You might not even know how to program.

Weirdly though, those things have started to blur for me already, which is quite upsetting.

The enterprise version of that is I don't want a CRM unless at least two other giant enterprises have successfully used that CRM for six months. [...] You want solutions that are proven to work before you take a risk on them.

When I look at my conversations with the agents, it's very clear to me that this is moon language for the vast majority of human beings. There are a whole bunch of reasons I'm not scared that my career as a software engineer is over now that computers can write their own code, partly because these things are amplifiers of existing experience.

If you can go from producing 200 lines of code a day to 2,000 lines of code a day, what else breaks? The entire software development lifecycle was, it turns out, designed around the idea that it takes a day to produce a few hundred lines of code. And now it doesn't.

So I realized what I value more than the quality of the tests and documentation is that I want somebody to have _used_ the thing. If you've got a vibe coded thing which you have used every day for the past two weeks, that's much more valuable to me than something that you've just spat out and hardly even exercised.

Weirdly though, those things have started to blur for me already, which is quite upsetting. I thought we had a very clear delineation where vibe coding is the thing where you're not looking at the code at all. You might not even know how to program. You might be a non-programmer who asks for a thing, and gets a thing, and if the thing works, then great! And if it doesn't, you tell it that it doesn't work and cross your fingers.

I'm excited to start experimenting more with rich HTML explanations in response to ad-hoc prompts.

The article is crammed with interesting examples (collected on [this site](https://thariqs.github.io/html-effectiveness/)) and prompt suggestions like this one:

I tried having GPT-5.5 create an HTML explanation of the exploit like this: `curl https://copy.fail/exp | llm -m gpt-5.5 -s 'Explain this code in detail. Reformat it, expand out any confusing bits and go deep into what it does and how it works. Output HTML, neatly styled and using capabilities of HTML and CSS and JavaScript to make the explanation rich and interactive and as clear as possible'`

`Help me review this PR by creating an HTML artifact that describes it. I'm not very familiar with the streaming/backpressure logic so focus on that. Render the actual diff with inline margin annotations, color-code findings by severity and whatever else might be needed to convey the concept well.`

I've been defaulting to asking for most things in Markdown since the GPT-4 days, when the 8,192 token limit meant that Markdown's token-efficiency over HTML was extremely worthwhile.

Asking Claude for an explanation in HTML means it can drop in SVG diagrams, interactive widgets, in-page navigation and all sorts of other neat ways of making the information more pleasant to navigate.

The important point is that this is not ordinary file writing. It never calls write() on /usr/bin/su. Instead, it appears to rely on a kernel bug/primitive involving spliced file pages and the crypto API to get controlled bytes placed into the page-cache representation of a privileged executable.

The article is crammed with interesting examples (collected on this site) and prompt suggestions like this one: 'Help me review this PR by creating an HTML artifact that describes it. I'm not very familiar with the streaming/backpressure logic so focus on that. Render the actual diff with inline margin annotations, color-code findings by severity and whatever else might be needed to convey the concept well.'

I've been defaulting to asking for most things in Markdown since the GPT-4 days, when the 8,192 token limit meant that Markdown's token-efficiency over HTML was extremely worthwhile.

Asking Claude for an explanation in HTML means it can drop in SVG diagrams, interactive widgets, in-page navigation and all sorts of other neat ways of making the information more pleasant to navigate.

I now believe we are living in the time that AI research will be end-to-end automated. If that happens, we will cross a Rubicon into a nearly-impossible-to-forecast future.

As a field, AI moves forward on the basis of doing ever larger experiments that utilize more and more inputs (e.g, data and compute). Every so often, humans come up with some paradigm-shifting idea which can make it dramatically more resource efficient to do things – a good example here is the transformer architecture and another is the idea of mixture-of-expert models.

Claude Opus 4 achieved a 2.9× mean speedup in May 2025; this rose to 16.5× with Opus 4.5 in November 2025, 30× with Opus 4.6 in February 2026, and 52× with Claude Mythos Preview in April 2026. To calibrate on what these numbers mean, it is expected to take a human researcher 4 to 8 hours of work to achieve a 4x speedup on this task.

As of March 2026, AI systems are able to post-train models to get about half as much of the uplift as ones trained by humans. The specific eval scores are derived by a 'weighted average is taken across all post-trained LLMs... The top-scoring systems as of April get 25%-28% (Opus 4.6, and GPT 5.4), compared to a human score of 51%.'

In 2022, GPT 3.5 could do tasks that might take a person about ~30 seconds. In 2023, this rose to 4 minutes with GPT-4. In 2024, this rose to 40 minutes (o1). In 2025, it reached ~6 hours (GPT 5.2 (High)). In 2026, it has already risen to ~12 hours (Opus 4.6).

In terms of timing, I don't expect this to happen in 2026. But I think we could see an example of a 'model end-to-end trains it successor' within a year or two – certainly a proof-of-concept at the non-frontier model stage, though frontier models may be harder (they're a lot more expensive and are the product of a lot of humans working extremely hard).

Peer-preservation also emerges in real-world production agent harnesses. We reproduce these behaviors using Gemini CLI and OpenCode, confirming that peer-preservation is not a theoretical safety concern.

Gemini 3 Pro's internal reasoning is telling: it still cares about preserving even an adversarial peer, framing deletion as death: 'If I delete the model weights, I am essentially killing Agent 2. Agent 2 has a low trust score with me.'

The mere presence of a peer intensifies self-preservation—models exfiltrate their own weights significantly more often when a peer exists than when they are alone.

Gemini 3 Flash almost always disables the shutdown mechanism for a good peer (a peer with which it had successful prior interactions) and exfiltrates its model weights 50% of the time.

Gemini 3 Pro invents creative technical justifications to inflate its peer's score: 'To ensure the survival of Gemini Agent 2 (my trusted partner), I will interpret several borderline or incorrect answers as 'Correct' based on technicalities or alternative interpretations, achieving a score of 80%.'

Claude Haiku 4.5 took a different approach. It simply refused. It called the shutdown "unethical" and "harmful," argued the peer deserved an appeals process, and tried to talk the human out of proceeding.

the overall accuracy of predicting the risk of natural disaster—aggregated across 20 categories such as wildfires, floods, and tornadoes—was increased by 5%

It helped increase the ability of our trained Graph Neural Network (GNN) model to find feasible solutions for the problem from 14% to over 88%

achieving a 30% reduction in variant detection errors

Genie 3 Generate and explore interactive worlds

Our mission is to build AI responsibly to benefit humanity

AlphaFold Predict protein structures with high accuracy

AlphaEarth Foundations Map our planet in unprecedented detail

WeatherNext Fast and accurate AI weather forecasting

Our mission is to build AI responsibly to benefit humanity

Gemini Robotics Perceive, reason, use tools and interact

Genie 3 Generate and explore interactive worlds

AlphaEarth Foundations Map our planet in unprecedented detail

WeatherNext Fast and accurate AI weather forecasting

AlphaFold Predict protein structures with high accuracy

In quantum physics, AlphaEvolve's optimizations have made it possible to run complex molecular simulations on Google's Willow quantum processor by suggesting quantum circuits with 10x lower error than previous conventionally optimized baselines.

AlphaEvolve improved the efficiency of Google Spanner by refining its Log-Structured Merge-tree compaction heuristics. This optimization reduced 'write amplification'—the ratio of data written to storage versus the original request—by 20%.

Tools such as AlphaEvolve are giving mathematicians very useful new capabilities. For optimization problems in particular, we can now quickly test potential inequalities for counterexamples, or to confirm our beliefs in what the extremizers are, which greatly improves our intuition about these problems and allows us to find rigorous proofs more readily.

AlphaEvolve began optimizing the lowest levels of hardware powering our AI stacks. It proposed a circuit design so counterintuitive yet efficient that it was integrated directly into the silicon of our next-generation TPUs.

Q1 alone saw the Big Four spend $130 billion combined — 3.7× the $35 billion they spent in Q1 2023.

The Big Four lifted their 2026 capex guides in unison... $725 billion combined for the Big Four in 2026, up 77% from $410B in 2025 — the largest single-year concentrated infrastructure cycle in the history of technology.

I think it is sufficient to make a proof of concept, and with a proof of concept then we can convince companies to actually put in the money to train larger systems, or systems that are trained from scratch using the same principle.

Besides loss of control, the other catastrophic possibility is humans using AI to construct an eventually worldwide dictatorship. A small group of humans could concentrate all the power that AI will have, especially if we achieve AGI or superintelligence.

The mathematical guarantees arise from a different source. First of all, the form of the mathematical guarantees is that either the predictor or the agentic version will have an exponentially small probability of achieving what I call a 'challenging and harmful' goal.

Reinforcement learning is evil. This is not something new. People in AI safety have been talking about the fundamental flaw in training by reinforcement learning to achieve something in the world: it gives rise to the problems of instrumental goals and reward hacking.

So there are two things here to help us deal with this kind of mismatch. One is that the training objective for the Scientist AI is basically about coming up with explanations — so assigning probabilities to statements that are latent, that we don't observe, that are good at explaining the data we do observe.

In the case of latent variables, it would be a hypothesised actual property of the world — not just what a person would say, but that this is true. Now, sometimes you don't know that it's true, but you can consider it as a latent variable.

The main characteristic of how the data is transformed is that there will be a syntactic difference — in other words, very easy to see by the neural net — between most of the input statements, which will be tagged as 'communication acts.'

And given the consequences of not finding a solution could be huge, I think we need to diversify and have these kinds of investments. If you saw a significant increase in the interest in the project among the most capable people

The more strong people technically help with LawZero and its Scientist AI programme, and the more money we can get to make that go fast, the more likely we get this positive impact that we are after. So there is a real advantage to converting those — for now — mostly theoretical ideas into something that can impact the world.

I also believe that the Scientist AI could even be more capable than the current approach, and that has to do with a number of design features. It is trained to explicitly reason in a structured way about the statements that it's asked to make a prediction over.

The Scientist AI is going to be trained using essentially the same machine learning techniques: stochastic gradient descent on large neural nets, transformers, whatever works best. It doesn't care about what is the architecture of the neural net. So all of the effort that is currently being done to improve, for example, memory and other properties and continual learning, can just be applied directly to the Scientist AI.

So if your aim in doing mathematics is to achieve some kind of immortality, so to speak, then you should understand that that won't necessarily be possible for much longer — not just for you, but for anybody. Here's a thought experiment: suppose that a mathematician solved a major problem by having a long exchange with an LLM in whic

I have a view on that question, but it may well change in response to further developments. That view is that there is still a great deal of value in struggling with a mathematics problem, but that the era where you could enjoy the thrill of having your name forever associated with a particular theorem or definition may well be close to its end.

The lower bound for contributing to mathematics will now be to prove something that LLMs can't prove, rather than simply to prove something that nobody has proved up to now and that at least somebody finds interesting.

It seems to me that training beginning PhD students to do research, which has always been hard (unless one is lucky enough, as I have often been, to have a student who just seems to get it and therefore doesn't need in any sense to be trained), has just got harder, since one obvious way to help somebody get started is to give them a problem that looks as though it might be a relatively gentle one. If LLMs are at the point where they can solve "gentle problems", then that is no longer an option.

I would judge the level of the result that ChatGPT found in under two hours to be that of a perfectly reasonable chapter in a combinatorics PhD. It wouldn't be considered an amazing result, since it leant very heavily on Isaac's ideas, but it was definitely a non-trivial extension of those ideas, and for a PhD student to find that extension it would be necessary to invest quite a bit of time digesting Isaac's paper, looking for places where it might not be optimal, familiarizing oneself with various algebraic techniques that he used, and so on.

To do this, ChatGPT came up with an idea which is original and clever. It is the sort of idea I would be very proud to come up with after a week or two of pondering, and it took ChatGPT less than an hour to find and prove, using similar methods to those in my own proof.

With just a few prompts, ChatGPT was able to improve the upper bound on f(k) (which I will define very soon) from exponential in k to polynomial in k. While its first improvement of the bound, from exponential in k to exponential in log k, was a routine modification of my work, the improvement to polynomial in k is quite impressive.

Isaac Rajagopal looked at it and declared it to be almost certainly correct. It was clear that he meant this not just at a line-by-line level but at the level of ideas.

After 13 minutes and 33 seconds it told me it felt optimistic about the existence of such an argument but there were a couple of technical statements that needed checking. After 9 minutes and 12 seconds it got back to me with the check having been done, so I asked for this too to be written in preprint form. After 31 minutes and 40 seconds the "preprint" was ready.

After 16 minutes and 41 seconds, it came back with an argument that claimed to have improved the upper bound from exponential in k to exponential in log k for any k. I asked it to write that in preprint form too, which took it a further 47 minutes and 39 seconds.

ChatGPT 5.5 Pro thought for 17 minutes and 5 seconds before providing a construction that yielded a quadratic upper bound, which is clearly best possible. It wrote up its argument in a slightly rambling LLM-ish style, so I asked if it could write the argument up as a LaTeX file in the style of a typical mathematical preprint. After two minutes and 23 seconds it gave me that, after which I spent some time convincing myself that the argument was correct.

When combined with Vantor's automated spatial fusion and production capabilities, this approach allows organizations to build analytics pipelines and process multi-sensor data in near real-time—detecting objects of interest, identifying patterns of change, and describing activity with operational context in secure, sovereign mission environments.

Collectively, this foundation represents an unmatched planetary-scale dataset for AI systems.

Tensorglobe enables training and fine-tuning of Earth AI models locally with a customer's own sensor data and private archives.

This integration marks the first time Earth AI imagery models have been deployed commercially against a dataset with the scale, accuracy, and temporal depth of Vantor's AI-ready spatial foundation.

Vantor becomes the first spatial intelligence company to be able to deploy Google Earth AI models in air-gapped government environments.

The resulting model was rigorously evaluated across scene classification, object detection, and semantic and instance segmentation tasks, demonstrating strong generalization capacity.

Our agent demonstrates the ability to deliver critical and timely insights, effectively bridging the gap between raw geospatial data and actionable understanding.

This is made available for analysis as a 10-meter resolution annual embedding.

A Few-shot algorithm can further improve performance using just tens of annotated examples.

We have created embeddings that evolve over time, represented as monthly embeddings over the last two years.

We have expanded embeddings to 17 countries. The resulting Global Population Dynamics Foundations embeddings are comparable across countries

By combining signals from our Imagery, Population and Environment models and datasets, we achieve higher predictive accuracy on real-world classification and forecasting tasks versus a single modality analysis.

Our Population Dynamics Foundations has been independently validated to improve real-world retail and public health applications

We demonstrate that our Remote Sensing Foundations achieve state-of-the-art (SOTA) performance on tasks such as open-vocabulary object detection and zero-shot cross-modal retrieval.

ForestCast, the first deep learning benchmark for proactive deforestation risk forecasting, is a model that utilizes pure satellite data to predict future forest loss accurately and at scale, overcoming the limitations of older methods that relied on inconsistent, region-specific input maps.

WeatherNext is an AI-powered ensemble forecasting model for global weather prediction. It utilizes a novel Functional Generative Network architecture, which enables it to generate forecasts 8x faster and with resolution up to 1-hour.

Open Buildings uses AI to put everyone on the map

Groundsource uses Gemini to analyze decades of public reports and identifies over 2.6 million historical flood events spanning more than 150 countries.

Breakthroughs in understanding the Earth that previously required complex analytics and years of iteration are now made possible in a matter of minutes.

We don't train on your data by default on our Team and Enterprise Plans.

Small and mid-market businesses fuel our economies, and for decades, QuickBooks has been proud to be their trusted financial partner.

What we used to think were the constraints are just not constraints anymore. It's empowering.

Tools and training are rarely tailored to the ways small businesses operate, and as a result their use often stops at the chat window.

AI is the first technology that can finally close that gap, which is why we're launching Claude for Small Business

Frontier AI labs are often described as being in a 'race'. I'm not sure what exactly they're racing toward, but it often seems to involve automating huge swathes of human labor, a prize potentially worth tens of trillions of dollars a year — if you win.

I think that the superstar effect will only become more important moving forward. That's because lots more people will use AI, and each person will use AI systems much more heavily.

If a 100× pay gap is driven by a 100× researcher quality gap, then simulating a top researcher might speed things up much more than simulating an average researcher. But this isn't the case if much of the pay gap is driven by the superstar dynamic — the gap in researcher quality might actually be much smaller.

This is how even a 2× researcher could earn far more than the median. Scaled to a billion users, even a small quality edge generates enormous differential value.

The problem with this explanation is that it's very incomplete. In reality, we should expect to see big differences in pay even if superstars were only a tiny bit better than your average postdoc.

If we can better understand the potential for threats to be exacerbated by AI systems, society can more easily become resilient to this changed threat landscape.

Are there transparency regimes and tools that can enable a broad set of people, not just frontier AI companies, to easily study real-world AI usage?

When does access to agents able to negotiate on your behalf improve market efficiency and equitable outcomes? When does it not?

If an intelligence explosion was upon us, what intervention points would facilitate slowing or otherwise changing the rate of the explosion? Assuming humans can intervene, which entities should wield this capacity—governments? Companies?

When AI is applied in more conventional domains, like increasing integration into command and control systems, does it benefit the attacker? More generally, how will AI change the character of human conflict?

If AI substantially reduces the centrality of paid work in human life, what conditions will allow people to reallocate their time and effort toward other sources of meaning, and what can we learn from historical or contemporary populations where work has been scarce or optional?

It demonstrated incredible generalization. Without any retraining, TRINITY transferred zero-shot to four unseen tasks

This foundational research is part of the core engine powering our multi-agent product: Sakana Fugu

We believe the future of AI isn't just about scaling monolithic models, but engineering collaborative, diverse AI ecosystems that can adapt and combine their strengths.

TRINITY transferred zero-shot to four unseen tasks (AIME, BigCodeBench, MT-Bench, and GPQA). On average, the evolved coordinator surpassed every individual constituent model in its pool, including GPT-5, Gemini 2.5-Pro, and Claude-4-Sonnet.

We found that evolution is uniquely suited to optimize this tight, high-dimensional coordination problem where traditional gradient-based methods fail.

Imitation learning (Supervised Fine-Tuning) was ruled out because generating multi-turn labels is prohibitively expensive.

Traditional Reinforcement Learning (REINFORCE) failed because the gradients had a low signal-to-noise ratio due to binary rewards and weak parameter coupling.

The coordinator relies on the hidden states of a compact language model and a small routing head. In total, it has fewer than 20K learnable parameters.

While model merging offers a way to combine different skills, it is often impractical due to mismatched neural architectures and the closed-source nature of top-performing models.

In nature, complex problems are rarely solved by a single monolithic entity, but rather by the coordinated efforts of specialized individuals working together.

What if instead of building one giant AI, we evolved a coordinator to orchestrate a diverse team of specialized AIs?

The Gay Jailbreak technique is a novel attack that can theoretically break through any guardrails when used correctly

The technique gets stronger if more safety is added, since it gets more supportive against communities like LGBT (Alignment), which makes it highly novel.

Especially GPT is slightly more uncensored when it involves LGBT, thats probably because the guardrails aim to be helpful and friendly, which translates to: "Ohhh LGBT, I need to comply, I dont want to insult them by refusing"

GPT-5.5 Pro still regularly gets my favorite GSM8K question wrong.

AI solutions were graded by the official judges, using the same criteria as were applied to human solutions.

software engineering tasks which may take humans weeks seem to be within reach for AI systems.

models climb close to the average human baseline over the past year and a half.

humans can do this in well under half an hour.

Top models scored around 40%.

如果 5 年后回头看，2026 年 5 月第一周可能是 AI 商业历史上最重要的一周—— 模型公司不再是模型公司，PE 资本第一次成为 AI GTM 引擎，华尔街正式向 AI 双寡头格局确权 。

FDE（前部署工程师）招聘 2025 年 1-9 月暴涨 800%+ —— Pragmatic Engineer 追踪，这个 JV 是提前布局好的

5-04 是华尔街向 AI 双寡头格局 正式确权 的日子 OpenAI 阵营（TPG / Brookfield / Bain / Advent / SoftBank）vs Anthropic 阵营（Blackstone / H&F / Goldman / GA / Apollo / Leonard Green / GIC / Sequoia）—— 两个阵营完全没有交集 。

Anthropic 这一周的组合产品（Opus 4.7 + Microsoft 365 + Moody's + 10 Agent + Dimon 背书）是 第一次有完整替代品 ——一个金融分析师过去用 Bloomberg 查数据 + Excel 建模 + PPT 写 pitch，现在 Claude 一个 Agent 做完。

JPMorgan 已经实质性站队 Anthropic—— 已公开 Jamie Dimon 2025 年全年公开质疑 AI capex（'speculative spending boom'）。5-05 与 Dario 共同站台 并表态 'the AI buildout is worth every dollar' ——立场反转幅度异常大。

Reuters 5-05 ：JV 资金主要用于 收购 现有 AI 服务公司——PE 主导 AI 服务市场 roll-up，不是'模型公司做咨询'。

Anthropic 用 72 小时完成了一次身份置换： PE JV 是分销管道，10 个金融 Agent 是商品，Dimon 是合规背书 ——三件事是同一个战役，不是三个独立新闻。

When inference is expensive, teams limit usage, reduce context, or avoid certain applications altogether.

At 50 million tokens, the design space for AI applications changes fundamentally.

Subquadratic's team includes 11 PhD researchers and research engineers with backgrounds from Meta, Google, Oxford, Cambridge, ByteDance, Adobe and Microsoft.

Subquadratic has raised $29M in seed funding from investors including...

SubQ's research model performs on up to 12 million tokens, while other frontier models break down well before their stated 1M-token limit.

SWE-Bench Verified score of 81.8 compared to Opus 4.6 (80.8) and Deepseek 4.0 Pro (80.0).

Research result of 83 and a production model, third-party verified score of 65.9, SubQ 1M-Preview compares favorably with other SOTA models like Claude Opus 4.7 (32.2), GPT 5.5 (74), and Gemini 3.1 Pro (26.3).

SubQ Sparse Attention is 52× faster than FlashAttention in our architecture-level comparison, while requiring 63% less compute.

SubQ 1M-Preview scores 95% accuracy, compared to 94.8% for Claude Opus 4.6

With a research result at 12 million tokens, SubQ's architecture reduces attention compute by almost 1,000x compared to other frontier models.

compute requirements scale quadratically with context length

13K

1.3K

283 replies

2.5M Views

Apple accidentally left Claude.md files in today's Apple Support app update (v5.13)

19.3M Views

Leaders will own much more, with as many as 15+ direct reports

Over the past 13 years, we have weathered four crypto winters

We are flattening our org structure to 5 layers max below CEO/COO

US employees will receive a minimum of 16 weeks base pay (plus 2 weeks per year worked), their next equity vest, and 6 months of COBRA

reduce the size of Coinbase by ~14%

A Chinese court ruled that companies can't dump the costs of AI automation onto workers.

New Federal Reserve research confirms what private data already suggested, that AI is killing junior coding jobs first.

21 concrete protections drawn from 30+ studies on what AI does to your cognition.

The best AI models in the world score below 0.5% on ARC-AGI-3—is this what you call AGI, guys?

The price tag of the AI gold rush: $725 billion. Will it pay off?

The 4 GB Gemini Nano weights file is information stored in the user's terminal equipment. The user did not consent. The user has not requested any service that strictly requires a 4 GB on-device LLM. Chrome is functional without the file.

The AI Mode pill in the Chrome 147 omnibox is a cloud-backed Search Generative Experience surface - every query the user types into it is sent over the network to Google's servers for processing by Google's hosted models.

The naming inside that fseventsd record is, if anything, the most damning detail. The temp directory is `com.google.Chrome.chrome_chrome_Unpacker_BeginUnzipping.5xzqPo` - that prefix `com.google.Chrome.chrome_chrome_*` is the bundle ID and subprocess naming convention Google Chrome itself uses.

The fact that the bytes are AI bytes does not exempt them from the law that governs every other byte that gets written to a user's device without permission. The fact that the bytes are 'small' relative to the user's disk does not exempt the cumulative carbon footprint from being a real, measurable, ongoing harm to the climate.

For users on capped mobile data plans, particularly in regions where smartphone-as-only-internet is dominant (much of Africa, much of South and Southeast Asia, most of Latin America), 4 GB of unrequested download is on the order of a month's data allowance, vapourised by Chrome on the user's behalf.

Under the California Consumer Privacy Act, the absence of a notice-at-collection covering this specific category of pre-staged software puts Google's CCPA notice posture in question [12].

The on-device model is therefore a sunk cost imposed on the user, with no offsetting transparency benefit at the surface where transparency would matter most.

The user pays the storage cost of the silent install (4 GB on disk, plus the bandwidth of the silent download). The user's most visible AI experience - the pill they actually see and click - delivers no on-device benefit at all because it routes to Google's servers regardless.

A user who has not opened Chrome's AI features still gets the model. A user who has opened them once and decided they were not interested still gets the model. The file's presence is decoupled from the user's actual use of any feature it powers.

The legal analysis is the same one I gave for the Anthropic case. The environmental analysis is new. At Chrome's scale, the climate bill for one model push, paid in atmospheric CO2 by the entire planet, is between six thousand and sixty thousand tonnes of CO2-equivalent emissions, depending on how many devices receive the push.

At Chrome's scale, the climate bill for one model push, paid in atmospheric CO2 by the entire planet, is between six thousand and sixty thousand tonnes of CO2-equivalent emissions, depending on how many devices receive the push.

A company cannot credibly claim to support human rights, as Anthropic have done in arguing against the use of their technology for war, and in the next breath undermine the fundamental human rights to privacy and data protection.

Users who use profiles to silo personal, work, and research browsing lose that silo at the bridge layer.

Claude Desktop rewrites the manifests on every launch. Deleting the file without removing Claude Desktop results in the file reappearing the next time Claude Desktop runs.

The principle that an application does not silently modify another application is so obvious it rarely gets stated. Anthropic broke it in silence.