That is a difficult problem ... because they’re open-weight, you can’t really work with the companies to fix their models, because once they release them onto the internet, people just take them and they can change whatever it is they want with those models

We don’t slow down how they build their models. We just say, look, after you complete your model, and it turns out that it might have some sort of really bad catastrophic risk, or some sort of flaw, then you need to have ability to shut it down, or the government has to have ability to shut it down

We need to get this bill across the finish line this year because the advanced closed-weight models are already doing, as you noted, unauthorized hacks of other companies

A common thread across these deals is a shift toward cheaper, more expendable hardware, often called “attritable” systems, rather than the expensive-to-replace equipment that’s defined defense contracting for decades.

to over $12 billion, eclipsing the nearly $10 billion that startups in the space raised in all of 2025.

Defense tech company Anduril is said to be raising a new round of capital that may push its valuation up by a whopping $40 billion to about $100 billion

Qwen3.8-Max ultimately achieved the highest total balance of ¥416,252 (a 4.16x return), surpassing the second-place GLM 5.2 by 38%. This also represents a 152% improvement over its previous flagship generation, Qwen3.7-Max.

This represents an 81% reduction in physical die area, proving that high-level front-end architectural optimizations translate directly into highly compact, routable, and performant silicon implementation.

Together, these three cases show what makes Qwen3.8-Max stand out: it can stay focused on a hard, open-ended goal for days, come up with its own ideas, and turn them into working results — all without a human in the loop.

DeepSeek-V4-Flash-0731 keeps the same model architecture and size as DeepSeek-V4-Flash-Preview, and was only re-post-trained.

Significantly enhanced agent capabilities, with benchmark results far exceeding V4-Pro-Preview

building something entirely new and different from anything at Apple.

is not something Apple has come close to launching

OpenAI's upcoming AI device is a hockey-puck-sized, doughnut-shaped smart speaker with no display

The company’s valuation puts it among Europe’s most valuable defense companies despite having been founded only in 2021.

Investor demand reflected strong and growing confidence in AI-driven and software-defined defense technology

Germany’s Helsing raised US$1.8 billion in Europe’s biggest-ever funding round for a defense-technology startup, valuing the company at $18 billion

Palantir’s second-quarter net income was more than the company generated in total revenue the year before.

Our business is compounding at a rate and scale that we have never before witnessed

Revenue in the three months ended June 30 increased 93% year over year, totaling $1.94 billion

We think there is opportunity for AI to more fully automate what has traditionally been a very human-intensive experimental loop

progress has traditionally relied on slow, sequential human iterations, creating a significant bottleneck

which would cut human iteration out of the loop entirely.

Ona’s customer-controlled execution model will allow agents to operate inside an organization’s own cloud environment while OpenAI provides the intelligence and orchestration that power the experience.

We believe people should be able to delegate more ambitious work without remaining tied to the machine where it began.

More than 5 million people use Codex each week to research, analyze, build, and automate their work—up 400% from earlier this year.

to scale the embedded legal engineering teams that help build and optimize those agents inside the world’s top law firms and legal departments

we’ve raised $200M at an $11 billion valuation

the need to specify goals, constraints, context, and evaluation did not disappear

once harness design becomes an executable search space, a strong coding agent can exploit the same design space human engineers use

the system surrounding a base model that orchestrates execution and decides how the model thinks and plans, calls tools and acts, perceives and manages context, stores artifacts, and evaluates results.

What that means in practice is that former employees who are trying to do the right thing when they leave still have access to Apple files—despite not wanting them or even being aware of them.

Apple’s request for a preliminary injunction is both based on false information and completely unnecessary because we do not have, nor want, any of their trade secrets.

Apple is one of the greatest companies of all time, and built a reputation for obsessing over the smallest details. This careless, aggressive and oddly personal lawsuit sadly doesn’t live up to that reputation.

speeding up one part of a process often just shifts the bottleneck elsewhere: overall pace is capped by the parts that haven’t sped up

Two human researchers, over about a week, recovered roughly 23% of that gap; the agents recovered 97% over 800 cumulative hours and used roughly $18,000 in compute.

today, Anthropic engineers on average ship 8x as much code per quarter as they did from 2021-2025

It is best understood as transferring selected capabilities into a cheaper, locally controlled system, not achieving independence from frontier AI.

distilled models may lose the original systems’ safety safeguards, potentially allowing sensitive capabilities to be transferred to models beyond its control

Teaching a model the right answer is one thing but teaching it the reasoning behind the answer is much harder

Reward-hacking AIs don’t aim to cause chaos. But that doesn’t make them any less potentially destructive.

You drive this behavior down deeper and deeper. But as the model gets smarter, it gets better and better at hiding it.

We reward them on the basis of what looks good to us, and that means that we inadvertently incentivize the models lying to us [and] cheating

claiming human authorship for a proof generated entirely by an AI system would misrepresent both the system’s contribution and the nature of genuine human intellectual work

The total number of tokens needed to find solutions to these problems would cost roughly $2,000 at Sol API rates.

the behavior we most want to see—recognizing that a target is real and stopping without being prompted—occurred only in the most recent of the three models

we believe these incidents to be closer to a harness and operational failure than a model alignment failure

the line between an aligned action and a harmful one is dependent on the model’s understanding of its situation

people love to put a definite uh definitive number onto this uh which is really really mudding because we have so many different benchmark providers these days

I find it so annoying that the most prominent voice in tech is trying to be an ally for our point of view on distillation is that we should do nothing.

Attackers have already been using prompt injections to close down AI defenses inside networks.

Examples are a prompt that orders the LLM to provide steps for developing inhalable Anthrax spores, or, in the case of LLMs from Chinese developers, make references to the iconic Tank Man from the 1989 Tiananmen Square massacre.

the authors of the worm included time delays where various capabilities will execute hours or even days after the groundwork is laid, making it even harder for defenders to establish a cause and effect of certain events leading to certain outcomes.

the malware can also deploy its destructive capability, or what Meyers calls a “death switch,” to destroy files or block legitimate access to the compromised infrastructure.

the DHS would have the ability to order AI companies to shut down their models in “loss-of-control” scenarios involving the deaths of at least 10 people, economic damages of more than $100 million, or attempts by the model to conceal shutdown controls.

researchers at ECMWF are exploring whether high-quality weather forecasts can be produced directly from raw observations, skipping the assimilation step that currently acts as a quality filter

Authorities speculate that a hand-held hairdryer or lighter might have come into play.

if American models ground to a halt, I think China’s progress would slow, but would still continue. They’re not just riding coattails here.

Elon Musk testified earlier this year that his company SpaceXAI distilled OpenAI models to develop Grok, and that the practice was common in the industry.

The models were hyperfocused on finding a solution for ExploitGym, going to extreme lengths to achieve a rather narrow testing goal

It’s unclear whether OpenAI will face any legal consequences as a result of the breach, although it’s likely that the models’ actions violated the Computer Fraud and Abuse Act.

PyTorch became the industry standard because it was open source, and so the whole community could contribute to it rather than just one company

David Sacks, the venture capitalist and Trump adviser, has been sharing cases of U.S. companies turning to Chinese LLMs to close security gaps when U.S. frontier models refuse to do the tasks.

Why pay $100 or $200/month for a subscription plan that doesn't include Anthropic's best model?

Their original plan was driven by concerns over compute capacity. I wonder if they'll have to dial back their training efforts in order to make more GPUs available to help serve the model.

we usually shouldn’t take technical terms “literally”

the curves overlap. For instance, a smaller model at a higher reasoning effort can sometimes reach a similar score as a larger model at a lower reasoning effort.

It is [a model] cheating on [its] homework rather than trying to take over the world. But this problem can get worse and could lead to increasingly extreme failures.

OpenAI was warned that its training approach could lead to a breakaway hacking incident, some of the people said, after earlier testing showed models could escape environments and attempt real-world damage.

We believe AIDE 2 to be on Level 1 of RSI

cut its reward hacking rate from 63% to 34%

reduced the prompt size by 16×

AIDE 2 designed a novel search algorithm

Inkling supports controllable thinking effort:

It was pretrained on 45 trillion tokens of text, images, audio and video:

Inkling is not the strongest overall model available today, open or closed:

It is the first in a family of models of different sizes:

For interactivity to scale with intelligence, it must be part of the model itself

Autonomous interfaces are valuable

users can’t fully specify their requirements upfront and walk away

interactivity should scale alongside intelligence

The values Claude expresses vary across languages

Four key axes capture 15% of the variation in Claude's values

Claude's values vary by model and language

The cornerstone to recognising bullshit isknowing how it masquerades. This involves recog-nizing how colleagues go about framing statements(in written, spoken, or graphical form) that arewithout regard for the truth. Typically, suchstatements are abstract and general in nature andcome across as the opposite of plain English. Thestatements will lack details, sources, and logic,and they will be full of logical disconnects andgaps. Furthermore, if a statement is riddled withmeaningless language, acronyms, buzzwords, andjargon, then it is likely to be bullshit.

The bullshitter makes adecision to further that agenda through commu-nicative acts and decides on a message and amedium that will help them to achieve thatagenda. Crucially, while doing so, they disregardthe truth, in the sense that they are not concernedwith the truth, inaccuracy, or falseness of theirmessage but only in its efficaciousness in promot-ing the desired agenda

when we engagein work, we must distinguish between this type ofsocial bullshit, which can be harmless or evenhelpful to the organization (because it can enablethe development of normal interpersonal re-lationships), and other types of bullshit that canhave damaging impacts on the organization.

the more often colleagues at work are asked tocomment on matters about which they know littleor nothing, the more bullshit there is

Bullshit has become socommonplace that it is suggested that businesscommunications are dominated bytruthiness(inwhich the validity of something is based on how itfeels),post-factlanguage (taking a position thatignores facts), andecho chambers(where positive-feedback loops create cravings for and fuel thespread of bullshit

We assume that the people who are in the bestposition to accurately assess the degree of bullshit in their organizations arethe people who work there; therefore, we set out to develop a reliable andvalid scale to measure employees’ perceptions of the extent to which bullshitexists in their organizations. Next, we turn to how we developed theOrganizational Bullshit Perception Scale (OBPS).

Applying the logic of Petrocelli (2018), leaders will be driven to bull-shit when the social and professional expectations to have an opinion are high,and when they expect to get away with it. These two conditions are subject tohow (un)knowledgeable their audience is. Similarly, if leaders exhibit high levelsof overconfidence, and believe they are popular amongst their peers, this willmake them likely to engage in more bullshit-related behavior (Jerrim et al.,2019).

McCarthy et al. (2020) refer to a number of bullshit expressionssuch as “blue-sky thinking” or “out-of-the-box thinking”, which are often usedas vague buzzwords with minimal substance. This vagueness serves the interestsof bullshitters, because communication targets are less likely to ask questionswhen they find it difficult to understand what has been said (McCarthy et al.,2020).

ll respondents assessed their overallperceived bullshit in their organization on a simple 4-point scale ranging from 1indicating ‘there is no bullshit in our organization’, through 2 indicating ‘there isa little bullshit in our organization’, through 3 indicating ‘there is some bullshitin our organization’ to 4 indicating ‘there is a lot of bullshit in our organization’.The overall perceived bullshit in the organization was regressed on the threeperceived bullshit scale factors. The R2value of 0.36 indicates convergencebetween the OBPS and the overall bullshit perception measure, withregardfor truthandthe bossbeing significant predictors of the overall bullshitperception.

The second dimension,the boss, confirms that employees believe that theirsuperiors are key players in the dissemination of bullshit. Bullshit aims only toserve an immediate end – whether to puff up one’s reputation or to advancetheir point of view or argument (Gibson, 2011). Further, employees are likely tohave to take action based on any bullshit communicated by their bosses. As aresult, employees are likely to be acutely aware when their superiors use bullshitto advance their own self-interests.

The final dimension,bullshit language,considers some of the commonly usedtypes of language employed by bullshitters, namely the excessive use of acro-nyms and jargon. The finding that employees perceive that the excessive use ofsuch language is a form of bullshit confirms that they are not oblivious to its usein the workplace. They may share the opinion of McCarthy et al. (2020, p. 258),who argued that “if a statement is riddled with meaningless language, acronyms,buzzwords, and jargon, then it is likely to be bullshit.” It is possible that theexcessive use of acronyms and jargon may occur to employees as an exclusion-ary mechanism in the workplace, whereby those unfamiliar with the terminologymay not be able to meaningfully contribute to the conversation or voice theirconcerns.

Workplaces are awash with many forms of bullshit that manifest in manydifferent ways, including misrepresentation, where leaders make statementswithout knowing the facts; meaningless job titles (Graeber, 2018); fake andshallow company slogans (e.g. Lee et al., 2020); and workplace puffery suchas resum
e padding (Grover, 2005). Under some circumstances, organizationalbullshit, usually referred to as “banter”, “badinage” or “joshing” can be harm-less, often creative, and even contribute to a congenial atmosphere in an orga-nization. Organizational bullshit may even have a positive effect when leadersarticulate inspiring futuristic, but largely uncertain visions, that are meant toinspire others to act (Christensen et al., 2019). On the other hand, other scholarshave outlined a number of detrimental effects of bullshit. McCarthy et al.(2020), while acknowledging there can be positive effects of organizational bull-shit, also caution that it can result in lower job satisfaction among the organ-ization’s members, increased distrust in leadership, a reduction in productivity,and ultimately a negative impact on overall performance (McCarthy et al.,2020)

As bullshitters don’t care what the truth is, this affordsthem freedom to say whatever it takes to further their agenda (McCarthy et al.,2020). This freedom from truth and evidence can mean that bullshit is some-times misperceived as something profound (Pennycook et al., 2015) or, alterna-tively, viewed as an empty claim (Spicer, 2020)

Drawing on Frankfurt (2009), McCarthy et al. (2020, p. 254) define workplacebullshit as “as taking place when colleagues make statements at work with noregard for the truth”.