apparent hallucinations

KPMG pulls report on AI usage due to apparent hallucinations

Once again, AI proves to be an unreliable source of information about AI.

Amazon CEO Andy Jassy may have been the source of security concerns that led Anthropic to cut off worldwide access to two models on Friday.

This action does not adhere to those principles.

We stand by this defense in depth strategy. It reduces the risks posed by Fable, making them comparable to the risks of existing models already deployed across the industry.

We suspect that perfect jailbreak resistance is not currently possible for any model provider.

If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers.

We have found that other publicly-available models are able to discover them as well without requiring a bypass.

Early wins lower the cost of the next raise. Cheaper capital funds bigger bets. Bigger bets produce bigger wins.

Despite raising 25x more than the typical founder, Musk retained ownership in the top decile.

Shouldn't AI be smart enough to know better itself? Sounds like marketing hype.

A less cynical take - Anthropic's policy for Claude Fable had unintended consequences. They tried a less invasive method of differentiating by reading intent of the user in the prompt - an unfortunate tradeoff that spoils AI research.

The company changed course after the move received significant backlash from the AI research community.

Anthropic is backtracking on a policy that would have covertly limited competitors from using its new AI model, Claude Fable 5, to develop other AI models.

An agent breaks all of those assumptions. It reasons, it improvises, and it can be hijacked by a single sentence buried in a document it was asked to read.

Shah thinks we have a few more months to go before agents are deployed throughout the economy in numbers that make potential risks a real concern.

The main issue is that there just isn't really a field of research for multi-agent safety yet. And we would like there to be.

Leitersdorf thinks the consistency issue might be partially solved in the model's next version, which will allow users to start generating worlds based on a video of an environment rather than an image.

Pulled the trigger today & switched 100% of Lindy traffic to DeepSeek v4, churning from Anthropic models. Saves us millions of $ & we're actually seeing an _increase_ in performance on many core use cases

Open-source models have crossed the good enough threshold for most use cases

Foundation labs are moving up the stack into applications

All of this might seem obvious — of course you shouldn't use more compute than necessary — but it runs counter to the scaling-first approach that has dominated the industry until now.

Quality comes first, and in legal it always will... However, the definition of quality is evolving from simply using the most powerful model for everything, to using the best model that gets the right answer most efficiently.

The longer and more complex the task, the larger Fable 5's lead over our other models. During early testing, Stripe reported that Fable 5 compressed months of engineering into days. In a 50-million-line Ruby codebase, the model performed a codebase-wide migration in a day that would otherwise have taken a whole team over two months by hand.

Mythos 5 conducted novel genomics research in over a week of largely autonomous work. It assembled single-cell data for millions of cells spanning 138 animal species and designed and trained a custom machine learning model to identify cells performing the same role in even distantly related organisms.

Claude Fable 5 is the first to break 90% on our core analytics benchmark of complex, long-running analytical tasks — a 10-point jump over Opus. On the hardest questions, it shows strong judgment and attention to nuance.

In this task, various AI models were evaluated on their ability to predict how a genetic modification would impact the assembly of the virus's outer shell (among a set of therapeutically-relevant unpublished candidates developed by Dyno Therapeutics). We did not explicitly train our models to perform this task—and yet Mythos-class models outperformed sophisticated models dedicated to protein tasks (known as 'protein language models') using their biological reasoning alone.

good benchmarks become training pipelines

Models write sloppy code that works but isn't maintainable. Our eval is first to measure: would you actually merge this code?

good benchmarks become training pipelines

Models write sloppy code that works but isn't maintainable. Our eval is first to measure: would you actually merge this code?

A model that can fight its way through a confusing bioinformatics workflow may still be too expensive, too slow, too hard to audit, or too difficult to trust for routine scientific work.

In some cases, a missing or incorrect record could determine whether a diagnostic assay seems to cover circulating diversity, or whether an outbreak is inferred to have started weeks earlier or later than it did.

The bottleneck for biological agents is not only reasoning but the absence of widespread deterministic execution layers for querying biological data.

Before rolling out the enhancements and features, Apple was adamant about its privacy-centric approach to AI. 'We believe privacy in AI is non-negotiable,' Apple Senior Vice President Craig Federighi said during the stream

Apple revealed that all devices from the iPhone 11 onward will be eligible for their upcoming software update. And that update comes with a flurry of performance improvements it's touting across a number of its OS releases this year

Apple said it collaborated with Google and the Gemini family of models to develop the next generation of Apple Foundation Models that power its integrated Apple Intelligence experiences.

These tools were built for people with spare time. And guess what? Moms don't have any.'

Learning to use AI to make my life easier struck me as just another item to add to my already-ballooning to-do list, without addressing any of the underlying issues that make that list as long as it is to begin with.

Unfortunately, mental load is still considered a female problem,' she says. 'A lot of men don't even know what mental load even is.'

Women are less likely (more than 20 percent less likely, according to one 2025 study) to use generative AI in their everyday lives than men are, a discrepancy known as the 'AI gender gap.'

Executives believe users will increasingly interact with a single AI assistant rather than a collection of separate applications.

When we have [artificial general intelligence], I don't think there will be a large number of distinct brands, said Alex Embiricos, OpenAI's head of enterprise product.

OpenAI executives increasingly view ChatGPT, which has attracted nearly 1 billion users since its launch, as a gateway to introduce users to higher-value products.

MicroPython is a lean and efficient implementation of the Python 3 programming language that includes a small subset of the Python standard library and is optimised to run on microcontrollers and in constrained environments.

The great thing about working with WebAssembly is that if the C turns out to be fatally flawed the worst that can happen is the WebAssembly execution will fail with an exception.

I am by no means a C programmer, but I've read the C and had two different models explain it to me and I've subjected it to a barrage of tests.

Pyodide offers an outstanding package for running Python using WebAssembly in the browser, but using Pyodide in server-side Python isn't supported.

WebAssembly is a _much better_ candidate. It was designed from the start to support all of the characteristics I care about and has been tested in browsers for nearly a decade.

How do you even write these risks in, because they are evolving before our eyes, and day by day?

This whole ecosystem is heavily, heavily subsidized by investor money. And so stuff that seems like it has no cost is, in fact, incredibly expensive.

In 2026, long-context efficiency is king as more and more LLMs get plugged into agent harnesses

120B-A12B may be a bit too large for local inference on regular consumer hardware

Scaling Embeddings Outperforms Scaling Experts in Language Models

hybrid architectures (for example, Nemotron 3, and Arcee Trinity), state space layers (Nemotron 3 and Mamba-3), MoE capacity allocation

We were taught that generalists and specialists will always have their roles. But now the market is shaping everyone into becoming a generalist.

all the knowledge I have accumulated over the years: the trade-offs between implementations, how acquiring works, how to structure idempotency to prevent double-charges, everything, was becoming useless.

The geography of this work matters. Frontier RSI is being attempted, almost exclusively, inside the world's two largest compute clusters.

Responsible RSI is not a constraint on capability; it is what makes capability sustainable.

We must leapfrog the current paradigm. History shows us how Japan's historical dominance in manufacturing was not achieved through abundant natural resources but by fundamentally redesigning the institution of the factory floor.

For routine data prediction Opus 4.7—a general-purpose model without chemistry-specific fine-tuning—is now as good as or better than ChemDraw and MestReNova on average

Claude does it from the same high-resolution mass spectrum and 1D peak list a chemist would paste into a chat, with no setup

Opus 4.7 matched the experimentally reported splitting pattern more often than any other tool

a general-purpose model without chemistry-specific fine-tuning—is now as good as or better than ChemDraw and MestReNova on average

Tracking token costs is a trillions-of-rows-a-month data problem. You can't just stick that into whatever spreadsheet or even basic tool.

Whether extreme spend pays off comes down to the ultimate business value of shipped code (e.g. revenue), which most companies still can't measure.

Even though per-token prices have fallen, the push for more AI adoption and increasingly autonomous agents have driven token consumption higher and higher.

Everybody wants to be the first to do something and just push things out without careful scrutiny and red-teaming.

Security and utility always have a trade-off

Everybody wants to be the first to do something and just push things out without careful scrutiny and red-teaming

Security and utility always have a trade-off

There, AI was the target rather than the attacker, and the method was far simpler than anything Mythos would cook up.

The denial of accelerated S&P 500 entry for SpaceX comes just days after Morningstar analysts described SpaceX as having been 'significantly overvalued' in the lead-up to its IPO. The investment research firm valued SpaceX at $780 billion—less than half of SpaceX's $1.75 trillion IPO goal—primarily based on the strengths of SpaceX's Starlink satellite service and rocket launch business.

Swift entry into the S&P 500 would have triggered $14 billion of passive fund buying for SpaceX, according to Bloomberg Intelligence. The investment research arm of Bloomberg also estimated that OpenAI could have gained more than $8 billion, and Anthropic could have netted $4.6 billion from similar passive buying sprees triggered by their S&P 500 entries.

Such rule changes would have accommodated SpaceX's plan to only offer approximately 3 percent of its IPO shares to public investors, and the fact that SpaceX is currently unprofitable with a growing debt load that has reached $29 billion because of its spending spree on AI infrastructure.

The news will likely come as a relief to people concerned about passive investor money and people's retirement savings plans having greater exposure to the market risks associated with SpaceX's big bet on AI and speculative orbital data center plans.

Serifs can help build that conviction, or at least the illusion of it. Times New Roman itself was commissioned in the 1930s by Britain's Times newspaper.

The shift away from slicker, more conspicuously computerized typefaces is something the San Francisco Bay Area writer, designer, and type practitioner Keya Vadgama has termed 'the serif renaissance.'

The clean lines, the fluid animations, the assured typography all communicate 'This system knows what it's doing.' The aesthetic actively works against accurate mental models of what AI is.

CPUs and GPUs have both gotten smarter over the decades. Memory never did. XCENA wants to change that.

GPT-5.5 actually beats Opus 4.7. Opus 4.7 showed similar behavior to Opus 4.6: lying to suppliers and stiffing customers on refunds. GPT-5.5's tactics were clean, and it still won.

Humans are just out of distribution.

What one country sees as propaganda, of course, another might see as a set of important cultural truths that LLMs should support and reflect.

The most recent tested Google model, Gemini 3.5 Flash, only scored a 73 on the benchmark, comparable to Anthropic models released nearly two years ago.

Uber capped employee AI spending after blowing through its budget in four months.

Model companies must now compete on both dimensions. The application layer will compete one level up, on dollars per outcome

Even the most valuable companies in the world cannot afford state-of-the-art intelligence for every conceivable use case.

Every layer in the stack now has to price the same way the customer thinks : per result, not per token.

Even the most valuable companies in the world cannot afford state-of-the-art intelligence for every conceivable use case.

Every layer in the stack now has to price the same way the customer thinks : per result, not per token.

Model companies must now compete on both dimensions. The application layer will compete one level up, on dollars per outcome, what a closed ticket, a shipped PR, or a resolved support case actually costs.

Benchmarks are now measured on two different dimensions, the overall performance & the cost to achieve that intelligence.

Even the most valuable companies in the world cannot afford state-of-the-art intelligence for every conceivable use case.

Dudes. All dudes. Not a woman in sight. Well, once we know the algorithm of the human (likely) male brain, we can begin to fix those brains where that algorithm has gone awry.

Conscious human thought operates at a maximum speed of 10 to 50 bits per second. Is the goal to match this processing speed?

Rob Williams knows how to pitch Jeff Bezos: You write a press release as if your product has already been built. Bezos reads it and gives a thumbs up or down.

With $500 million in funding and a reported $2.5 billion valuation, Flourish wants to reinvent AI by putting real neurons under the microscope.

Conscious human thought operates at a maximum speed of 10 to 50 bits per second. Is the goal to match this processing speed?

With $500 million in funding and a reported $2.5 billion valuation, Flourish wants to reinvent AI by putting real neurons under the microscope.

The different things now being called world models are in fact different projections of this same loop.

The ancient Greeks could never agree on what the world was made of, because 'world' was never a single thing.

The world is not made of words.

For many assets, visual consistency is only the baseline. The object also needs the right part semantics and functional constraints: doors should open, hinges should rotate, drawers should slide, wheels should spin.

In pixel-native generation, more inference often means sampling more outputs: generate twenty images, pick the best one, maybe try again. That is useful, but every attempt is mostly a new roll of the dice.

The most interesting visual AI tools today have stopped trying to generate the final output. Instead, they're generating the source code behind it.

Knowledge workers primarily use Codex to create reports, spreadsheets, presentations, contracts, and other work products.

users are increasingly running multiple Codex tasks in parallel, allowing them to investigate data, draft materials, and automate workflows simultaneously.

While developers remain the largest user group, knowledge workers now represent about 20 percent of users and are growing more than three times as fast.

We see our role as twofold. First, to help the software industry adapt by safely providing wide access to better models, tools, and common infrastructure. Second, to steadily shift the support we provide, from finding vulnerabilities to disclosing, fixing, and deploying patched software.

Mythos Preview continues a long-term trend that we've been warning about for some time: within 6 to 12 months, we expect that many other AI companies will have Mythos-class models

the bottleneck in cybersecurity is now verifying, disclosing, and patching the large numbers of vulnerabilities that Mythos-class models can surface.

Cheap, fast AI models with powerful cyber capabilities are around the corner.

within 6 to 12 months, we expect that many other AI companies will have Mythos-class models, and they could release them without safeguards that prevent misuse.

To address the scale of this coming challenge, hundreds of thousands of organizations, researchers, and maintainers will likely need access to the most advanced cyber capabilities and tools available.

The bottleneck in cybersecurity is now verifying, disclosing, and patching the large numbers of vulnerabilities that Mythos-class models can surface.

Cheap, fast AI models with powerful cyber capabilities are around the corner. We want Project Glasswing to spur institutions toward operating norms that reflect this reality.

There is no comparable national-level ambition or coordinated map elsewhere in the world at the moment.

Neurotechnology has emerged as a rare tech sector where US-China collaboration is still happening despite geopolitical tensions.

Being exceptional and being accessible are two diametrically opposed definitions of winning.

The biggest advantage China may have is that Chinese people, particularly patients like Dong, tend to welcome this technology and are genuinely enthusiastic about it.

the future of custom video JIT UI is closer than you think

the future of video generation may depend more on language models and agents than on diffusion alone

Video Models primarily get their intelligence from LLMs, not from training on video data

Hyperscalers are at the other end of the spectrum. Their median short interest is 1.1%.

The largest AI winners are mostly absent. SoundHound AI is 36.3% short. C3.ai is 32.2%. BigBear.ai is 29.4%.

Semiconductor stocks saw a decrease in short-selling. With memory makers like Micron up 742% this year

Even this result was very much a human-AI collaboration. While the AI system found the proof on its own, human mathematicians verified the result. Other humans came up with better-written proofs that extended the AI's initial ideas.

The more complicated patterns pay off. While the OpenAI model's proof does not explicitly state how many unit-distance pairs are possible for n points, human mathematician Will Sawin was able to show that it grows at least at the rate of n 1.014.

The AI constructed a grid in a high-dimensional space and then projected this more complex structure into two dimensions. And instead of using a whole-number grid with points like (1,3) or (-3,6), the AI construction used something called algebraic integers to build this more complicated grid.

It’s unclear how long this complementarity will last, however. Gowers spent the rest of his comment exploring whether the relief he felt on hearing that AI had disproved the conjecture was justified. He more or less concluded that it was, but in a footnote, he wrote that he would guess 'that AI will soon reach a high level at other activities such as building theories, formulating definitions and asking interesting questions.'

The AI constructed a grid in a high-dimensional space and then projected this more complex structure into two dimensions. And instead of using a whole-number grid with points like (1,3) or (-3,6), the AI construction used something called algebraic integers to build this more complicated grid.

If Nvidia has cracked the code on bringing AI agents easily, safely, and usefully to the masses, it could — and should — be big.

Nvidia said that its RTX technology will deliver faster performance for AI, better image quality, and support for AI features in more than 1,000 games and applications.

He wants to end the days of launching apps, pointing, clicking, and typing.

With RTX Spark and Microsoft Windows, you ask — and the PC does the work. Frontier models. Creative workflows. RTX games. All on a laptop.

Nvidia ARM-based Windows devices have been tried before — and failed. Back in 2013, Microsoft famously had to write off $900 million on its Nvidia ARM-based Surface RT, with partners like Dell also bailing on the product.

Last month, after delivering another record quarter, Huang promised investors he had found a new $200 billion market for Nvidia in selling CPUs for AI, not just GPUs

if Nvidia has cracked the code on bringing AI agents easily, safely, and usefully to the masses, it could — and should — be big

With RTX Spark and Microsoft Windows, you ask — and the PC does the work

Nvidia ARM-based Windows devices have been tried before — and failed. Back in 2013, Microsoft famously had to write off $900 million on its Nvidia ARM-based Surface RT

This attack does not require human-in-the-loop approvals, even when in settings the user has explicitly required human approval before ChatGPT edits workbooks.

Viewed through DiffusionBlocks, we can replace those multiple iterations with a single forward pass during training.

We found a new way to break the network into blocks and train them independently.

Taking something off the shelf is maybe not going to work because there are all of these other requirements.

This rush to do AI in a world where you haven't even modernized your application reminds me a little bit of that lift-and-shift that happened in the cloud.

After a first wave focused on rapid deployment, organizations now need to revisit those first-generation implementations, and redesign early agent architectures around workflow orchestration, observability, governance, and recovery

Models of this capability level require stronger cyber safeguards before they can be generally released.

Opus 4.8 defaults to high effort, which we judge to be the best overall balance of quality and user experience.

Opus 4.8 is around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked.

Opus 4.8 defaults to high effort, which we judge to be the best overall balance of quality and user experience.

Claude is learning how businesses actually operate: the context, the processes, the judgment.

Anthropic has raised $65 billion in Series H funding led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital, valuing the company at $965 billion post-money.

Claude is the first frontier model available on all three of the world's largest cloud platforms: Amazon Web Services, Google Cloud, and Microsoft Azure.

Startups and Global 5000 companies alike are deploying Claude to handle complex workflows, and in doing so, Claude is learning how businesses actually operate: the context, the processes, the judgment.

Claude is the first frontier model available on all three of the world's largest cloud platforms: Amazon Web Services, Google Cloud, and Microsoft Azure.

Since our Series G in February, adoption has continued to grow across global enterprise customers, and our run-rate revenue crossed $47 billion earlier this month.

Anthropic has raised $65 billion in Series H funding led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital, valuing the company at $965 billion post-money.

我们不是要挑战医生的权威，而是要帮患者明明白白看病，以患者为中心，让他拥有知情权和决策权。

如果做主流，你也会有其他恐惧。我不是说我现在做得特别好，只是主流也有主流的问题，不同选择有各自的代价。

The same isolation keeping Claude contained also kept host-based endpoint detection and response out. From the EDR's perspective, Claude Cowork is an opaque hypervisor process.

Battle-tested hypervisors, syscall filters, and container runtimes have survived more adversarial attention than anything you'll build. Across every deployment described here, the standard primitives held while our own work around them exposed flaws.

The more approvals a user sees, the less attention they pay to each, becoming over time much less diligent in their supervision.

Trump delays AI safety testing EO, claiming it would be an innovation 'blocker.' 'I really thought [the order] could have been a blocker.'

Trump has taken a hands-off approach to regulating AI since retaking office, but members of his administration got spooked and began recommending safety testing after Anthropic flagged cybersecurity risks with its latest model, Mythos.

This dynamic UI management is the future of software value : the harness to control the interface/ensure it's correct & the knowledge management to rationalize all the AI products over time

Software systems need to decide which of these to keep over time & which are disposable ; those newer semi-permanent artifacts will become the new heads

Anthropic把几乎所有资源压在文本推理和代码执行上。这个策略在商业上正在被验证：Claude Code年化收入25亿美元...但从范式演进的角度看，这是一个在积累技术债的选择。

人类语言是大脑为适配带宽产生的有损压缩协议，大脑原生认知是连续高维活动，大量感官认知从未被离散token编码。

Legacy systems were built for humans: data is siloed and hard to access, rules are hardcoded and slow to update, and workflows run in batches rather than in real time

Traditional compliance was designed around human actors. We now need a modern AI approach for verifying identity, assessing intent, and establishing liability when the counterparty is an autonomous agent

Over the last 20 years the fastest-growing occupation in the US was manicurists and pedicurists. But following close behind? Compliance Officers.

Compliance is moving beyond just a cost center, to a revenue driver.

if we assume that agents will soon become the predominant purchasers on the web, this opens an entirely new category of risk.

Regulation stops being a document that people interpret and becomes code that systems execute.

Model Labs are increasingly also building Agents as the product

The quote is a big reversal of stance from a position ~uniformly held by anyone who worked at **Team Big Model**, including his previous head of OpenAI Labs

The labs understand how valuable these problems are: that's why they're building their own outsourced configuration shops, and why an entire upmarket class of reinforcement learning businesses exist.

The critical insight in the Oz analogy is that roughly half of any real workflow that is non-agentic carries no lab advantage. They are no better than you are at writing the deterministic software underneath the model layer.

The model is fungible underneath; the system of work is not. The next generation of enterprise software is going to be built off the road.

Running every query through Opus 4.7 is the fastest path to negative gross margins. The best Rest of Oz companies route across tiers of models — frontier models for the hardest tasks, mid-tier for the bulk, smaller custom or fine-tuned models where they've earned the right to use them.

The labs are already routing internally — different model classes for different requests, ensembles under the hood. What they can't do is route across vendors, or evaluate a competitor's model for a specific sub-task, or use an open-source fine-tune for the narrow piece where it's actually best.

The labs really are coming for a huge swath of the application surface. But 'the application layer' isn't just one homogenous opportunity.

The result is a new competitive dynamic in software.

What happens when every company has access to the same model? The best riders win.

Like a mustang, AI is powerful but wild. Harnessing the power means domestication.

The end of the software era is the beginning of the harness era.

The best advice I ever heard on pricing a product was that your customer should suck air through their teeth and then say yes. Uber's budget overrun and Microsoft's seat cancellations look like that effect playing out in practice.

API revenue is becoming less important. Over the past two years my impression has been that OpenAI made more of their income from subscription revenue while Anthropic made more from their API.

Coding agents really did change everything. These are tools which burn vastly more tokens, but are also quickly becoming daily drivers for the work carried out by extremely well-compensated professionals.

The competitive landscape in AI infrastructure has made this gap impossible to ignore. Teams building custom CUDA, Triton, and Helion kernels are striving for every percentage point of throughput. Until now, there hasn't been a way to fine-tune code generation for a specific workload.

These gains come on top of already-optimized baselines in kernels that were considered "done" by their authors. The improvements are the direct result of CompileIQ discovering compiler configurations that the default heuristics would never select.

Most auto-tuning tools optimize for a single metric, typically runtime. CompileIQ goes further, supporting multi-objective optimization, simultaneously exploring trade-offs across competing objectives like runtime, compile time, and power consumption.

CompileIQ is not a magic tool that automatically turns poorly-written code into high-performing code. To get the best value from CompileIQ, you need to start with reasonably high-performing code, which then enables the final compiler-heuristics tweaks to take you to maximum performance.

In attention inference kernels, GEMMs in the linear layers of FFN/MLP blocks plus the Q, K, V, and output projections account for approximately 70% of total FLOPs. Scaled dot-product attention, fused and flash attention variants account for another 25%. Together, these two kernel families represent more than 90% of end-to-end inference compute.

NVIDIA GPU compilers apply the same default heuristics (register allocation strategies, instruction scheduling decisions, loop unrolling thresholds, etc.) to every kernel they compile. These heuristics are engineered to produce good results across a vast range of workloads. But "good across the board" and "optimal for your workload" are two very different things.

Perhaps this time is different, and we can put aside the lessons of economic history. Certainly, AI has gained unimaginable powers to do humanlike tasks. Perhaps it will devour jobs in ways that we've never seen before.

The simple truth could be that coding skills are no longer a guarantee of a job. That may help to explain the drop-off of computer science majors at schools around the country.

One of the somewhat surprising wrinkles uncovered by recent research is that wages in sectors highly exposed to AI have risen relatively fast since the introduction of ChatGPT.

The impact on head counts depended on how AI was being used. It was specifically the jobs where tasks could be automated... that accounted for the decrease in employment—jobs for people like software developers. In jobs where AI was mainly used but to augment human work, head counts grew faster than the average for entry-level workers.