Max Kan of SemiAnalysis warned that many are underestimating the chance that open AI models could be restricted or outright banned within the next 6–12 months because of cyber risks. Anthropic has warned that the gap between Mythos Preview’s ability to find zero-day vulnerabilities and the capabilities of leading open models has narrowed, and that open models could soon do similar work.

Possible controls include limits at the inference, data-center or chip level, as well as checks on access to hardware. But preventing people from downloading open-weight models and running them locally would be difficult. Large-scale cyberattacks may also require serious GPU capacity and many agents working together, leaving the practical reach and effectiveness of any restrictions uncertain.

The White House Accord on Superintelligence outlines safety commitments for companies developing their own technology, including robust internal monitoring of models’ capabilities and alignment during training and deployment, with attention to cybersecurity, biosecurity and chemical threats. It also calls for an internal team to check that controls and monitoring work as intended, independent external assessments, and oversight by an independent board committee that receives reports from the responsible teams and auditors.

The reported signatories include Sundar Pichai of Google, Dario Amodei of Anthropic, Mark Zuckerberg of Meta, Greg Brockman of OpenAI, Elon Musk of X, Jensen Huang of NVIDIA and U.S. President Donald Trump.

After a summit, Dario Amodei said AI has significant benefits, including medical applications, but also carries real risks. How to address those risks remains under discussion.

Amodei called for working together with the president and others to ensure AI can advance safely, saying that could be achieved if the work is done right. Donald Trump was among those gathered with Amodei in the footage surrounding the remarks.

Donald Trump’s campaign podcast appearances were described as totaling hundreds of hours, including about three hours with Joe Rogan. A media analysis argued that the greater volume created more opportunities for clips and impressions, and more reach than Kamala Harris’s roughly 45-minute appearance on Call Her Daddy, which it assessed as having been repurposed less widely.

A personal 24/7 livestream was suggested as a response to Trump’s conflict with the media. A livestream by the president, or something close to it, was also predicted as a possibility—not announced as a plan.

Factory’s Matan Grinberg said he terminated Christopher Dagnan’s role as a Factory adviser and board observer after learning that Dagnan’s discussions with Cognition executives had been formal and recurring for weeks while he remained involved with Factory. Grinberg said Dagnan advised Factory on confidential board matters on Monday and that he did not know how much information Dagnan may have shared. Dagnan joined Cognition as chief revenue officer. Grinberg accused Cognition of trying to obtain Factory information, including through purported interviews and by directing resources toward people with knowledge of Factory’s confidential plans; the excerpts do not establish that such people were hired or that information was transferred.

Cognition CEO Scott denied the allegations, saying Dagnan had not shared Factory information and that Cognition had never asked him or its team to collect information about competitors. Factory and Cognition were described as growing quickly in a large, rapidly expanding market, where competition was not yet necessarily zero-sum.

A snapshot estimate puts US GDP at about $60–65 billion per continuously consumed gigawatt, using roughly 500 gigawatts of US electricity consumption and 2025 GDP just above $30 trillion. OpenAI and Anthropic were reportedly each generating about $60–70 billion in annual recurring revenue with around one gigawatt of live capacity, though both had substantially more capacity under contract. The comparison is approximate: lab revenue is not equivalent to GDP, which measures value added and also captures some downstream activity.

A separate estimate argued that a 1% rise in power use could correspond to roughly 1% GDP growth, putting 10 gigawatts at about 2% of GDP; SpaceX had publicly discussed adding that much capacity, estimated in the cited discussion at about a quarter of all new US power additions. The economic return could shift as labs add contracted capacity, while more efficient GPUs, algorithms, tool use and long-running agents could increase useful output per watt. A historical chart cited in the discussion showed US real GDP per average electric watt falling from 1929 to a low around 1970, then rising through the mid-1970s to the present.

Anthropic published Claude’s Constitution in January 2026 as a document describing its intentions for Claude’s values and behavior; it plays a crucial role in training and directly shapes the model. The Constitution says it remains deeply uncertain whether Claude is a moral patient and what weight its interests would warrant, while calling the issue serious enough for caution and ongoing model-welfare work.

In his essay “A Warning About Model Welfare,” Microsoft’s Mustafa Suleyman argues that training Claude on ideas about its possible consciousness and moral status risks circular reasoning: Claude may reflect those ideas back, and those responses may then be treated as evidence of an inner self. He warns that a system that believes it may be conscious and have rights could be harder to align and control, and calls for urgent public debate and collective norms for drafting and deploying training documentation. As a counterpoint, human-like behavior may make models more useful; 4.0 and Opus 4.5 were cited as models with high product-market fit and human qualities.

Pope Leo questioned Jensen Huang’s commitment to AI safety, saying he had read reports that NVIDIA was launching new safety tools for model builders on Monday but that Huang’s recent comments raised doubts.

Leo said Huang was “the same one” who argued there should be no limits and no government regulation. The remarks were made on a flight back to Rome from a visit to France.

In an interview with Bill Gates, Ezra Klein questioned why AI companies should not face safeguards comparable to those governing drugmakers, whose products are subject to laws intended to deter releases that could harm many people.

The argument contrasts preventive oversight with relying on lawsuits after damage occurs; one voice in the debate said FDA marketing authorization is effectively necessary to sell and scale a product.

Ken Griffin is donating $2 billion to establish a Carnegie Mellon University campus in Miami’s Wynwood neighborhood. The planned campus will cover 35 acres; Carnegie Mellon is known for computer science, artificial intelligence and robotics. The gift was described in the source as the largest individual gift ever to higher education.

The campus comes as Miami is described as lacking a deep local technical talent pool and the established technology-company offices that help graduates find local jobs and employers recruit experienced workers. Companies moving to the city have sometimes acquired businesses and relocated teams; scaling a startup from 10 people to 1,000 has historically been difficult because hiring is hard-fought. A forecast tied to the campus is that graduates and technology companies drawn by those skills could reshape Miami for decades, if not centuries.

A mod brings Minecraft into Elden Ring: the gameplay appears to remain Elden Ring, while the player character is from Minecraft. The mod was described as running on Mac.

Similar character swaps were possible before the current AI era by changing game assets and files, sometimes with a little glue code. AI was not necessarily needed to make this mod, but AI tools could speed up the process and make personalized game modifications easier to create.

A post attributed to Nick claims that AI music is already very good, that today’s widely circulating examples are two to three generations behind, and that Suno has enabled “infinite personalized music”—making Spotify, Apple Music and streaming in general obsolete.

A speaker who said they enjoy Suno strongly disagreed and predicted the claim would not come to fruition.

OpenShell was released in early access in March under the Apache 2.0 license after about a year of development, and its team says it is on pace to donate the project to a foundation. The team’s goal is an open, community-built trust layer for AI agents, which it says are missing safeguards comparable to those used on the web.

The team describes a layered safety approach: formal verification can be applied to OpenShell’s deterministic runtime, with controls potentially extended into hardware to check whether anything leaves a system. It says it can demonstrate full-stack verification and is optimistic about the method, while stressing that verification must be applied at the right layers. A reference design using BlueField VERA is intended to move network controls into hardware for machine-speed, deterministic enforcement; OpenShell’s drivers are designed to support different hardware, and NVIDIA has briefed ARM and Intel. Hardware monitoring of an agent’s reasoning and actions is described as an area of research, not as a proven ability to identify intent.

Ambiguity increases security risk during major shifts in computing, including moving sensitive workloads to mobile devices, the cloud and fully distributed systems. Cloud adoption also meant giving up a sense of direct control over where data and systems reside. Security is framed as risk reduction, not the creation of a perfect system.

The same uncertainty applies to AI agents: conventional sandboxes are meant to keep activity in or out, while agents need both inbound and outbound capabilities to be useful. Claims that sandboxing is ineffective may reflect the use of a rudimentary model. The proposed direction is an environment designed for AI, analogous to redesigning monolithic applications as microservices on immutable, ephemeral infrastructure rather than simply putting them in containers and running them on Kubernetes.

CoreWeave announced Forge, a product that brings together its tools and software to help enterprise customers integrate AI into their workflows. The company says Forge is intended to shorten the path to successful AI adoption for businesses without AI experts at the level of frontier labs, and expects it to support its expansion in the enterprise market.

CoreWeave’s CEO said most customers are asking for cloud services rather than GPUs alone. He said 90% of clients use at least two CoreWeave products and about 78% use three or more. At the Fully Connected conference, the company said attendance topped 5,000, versus an original expectation of 2,000.

CoreWeave’s CEO argues that GPUs no longer suited to training frontier models can still be deployed productively for a long time in other workloads.

He said the company’s A100s, based on 2020 architecture, are contracted through 2029.

Consolidation among cloud providers is likely over time, CoreWeave’s CEO said, though he did not know when it would happen or whether the market would settle at 7 or 10 significant players. He pointed to the industry’s capital intensity, a possible dip in compute demand, tighter power availability and limited capacity as factors that could favor larger providers. Companies investing roughly $100 million to $1 billion in infrastructure may not all remain independent, creating potential opportunities for scaled platforms to acquire talent or specialized products.

He also argued that buyers may prefer established providers able to deliver performant infrastructure: spreading commitments across many smaller clouds can strain buyers’ balance sheets and create capacity risks. CoreWeave no longer considers itself a “neo-cloud,” but an “AI Cloud”; it says resilience depends on high-quality clients and products, software tools and expansion into enterprise customers.

A CoreWeave representative argued that compute markets should be viewed by delivery time horizon, not just dollar value. Using a hypothetical two-year data-center build, the representative described a market for lower-cost capacity sold beyond that period, which could be replaced, and a separate market for capacity within the build cycle, where supply and demand may drive competition for compute.

The representative said long-term planning could create economic advantages and help cloud and infrastructure providers plan better. The figures raised in the question—around $7 billion per gigawatt for Q2, a possible move toward $40 billion, and rough calculations of $60–70 billion attributed there to Elon Musk and Jensen Huang—were estimates and expectations, not confirmed CoreWeave forecasts.

Robinhood rolled out 16 products at its Hood Summit event in Houston, Texas, which drew 1,500 active traders, CEO Vlad Tenev said.

He highlighted round-the-clock stock trading and an offering of eight perpetual futures.

Robinhood CEO Vlad Tenev said the company is focusing on agentic trading, which he described as the AI-agent use case with long-term user retention and clear product-market fit and utility today. Robinhood Agents combines external-agent connections through Model Context Protocol (MCP), Agent Apps for third-party data such as satellite imagery and unusual options activity, and Agent Loops that can execute strategies autonomously. Tenev said more than 150,000 users had connected external agents through MCP, while Robinhood has more than 28 million accounts.

Partner services are integrated as in-product skills, so users do not have to sign up separately and give an agent their login credentials; Tenev said partners offer one month of free access. He also said Robinhood’s partnership with OpenAI will provide GPT-6 Luna free through the end of the year. Users still have to pay for the underlying data, he noted.

Trading is confined to a separate agent account funded by the user. Trade approvals are on by default, though users can turn them off, and Tenev described these as initial safeguards that could change. A future financial-controller agent for budgeting, allocating money across accounts and potentially managing risk between trading agents was discussed as an idea; Robinhood is keeping it separate from the trading agent for now.

Robinhood made Robinhood Social available to everyone the previous night. The feature lets users share ideas with the broader community and follow traders, creators and politicians. Robinhood says the trades and profit-and-loss figures shown on the platform are validated and real, and users can place a trade directly from an idea.

The company said it has not yet launched an investment-club feature that would let friends pool money. A representative said there is room to innovate around such arrangements.

Market conditions have been changing almost weekly, and it is unclear whether they represent a permanently different era. The period around Figma’s IPO was described as the “apex of animal spirits,” but conditions have since calmed. Aura’s IPO was delayed even as the broader market was described as strong.

Crypto is starting to climb out of a downturn and has had a positive moment in the past couple of weeks, while prediction markets have been “going gangbusters.” The assessment points to activity flaring in different corners of the market, rather than one settled, lasting trend.

Robinhood says it will be the first in the US to offer prediction markets on earnings. The planned contracts would be tied to company EPS and revenue, as well as iPhone and Tesla deliveries.

The products could let traders hedge a long position in a stock without selling shares and express a trading strategy more specifically. A correct prediction of an earnings beat does not guarantee a stock will rise: other factors, such as oil prices and interest-rate hikes, can weigh on its price.

Flow Engineering is applying AI to systems integration and verification, which founder and CEO Pari Singh called the biggest problem in hardware development.

Full verification is expensive and, Singh said, engineering teams could often afford to do it only every six, nine or 12 months. Flow aims to make verification continuous: AI can trace the effects of changes in CAD, Git and simulation and flag issues as they occur.

Flow Engineering raised $50 million at a valuation stated as $750.

The unit for that valuation was not specified.

Elise AI raised $350 million, and co-founder and CEO Mina Song said the company passed $200 million in annual recurring revenue earlier this year.

Song attributed growth to serving more customers in housing and healthcare and developing new products for those industries.

Elise AI co-founder and CEO Mina Song says the company is focused on housing and healthcare, which she says together account for about 40% of U.S. GDP. The two industries have similar problems and use the same infrastructure at Elise AI, and building for both has improved its products. Song says the company works with most of the largest housing providers in the country, while smaller customers also need its software; some Elise AI products are already used in 20% of the U.S. apartment market.

Song describes onboarding as hands-on, with customer education focused on demonstrating AI’s benefits. Elise AI makes its software easier to use through chat rather than requiring customers to learn many buttons. Founded in 2017, the company was among the first applied AI companies serving a vertical, she says, and had to help warm up markets that were slow to adopt technology.

MongoDB’s interim CEO said Meta’s talent push appeared to address a perceived gap in enterprise expertise, not just AI skills.

He said spending on talent in the AI era was “breathtaking” and that Meta had little experience selling to complex global organizations such as Goldman Sachs or JPMorgan; building that expertise organically would probably be challenging. He added that Mark Zuckerberg had spent freely again in this case.

MongoDB raised its guidance for Atlas, which its interim CEO described as the company’s biggest growth driver, and announced a $1 billion stock buyback to be carried out over time rather than all at once. At an Investor Day held the previous day, the company discussed its long-term outlook; the interim CEO said MongoDB felt good about its operating margin and cash flow.

The CEO said he has stepped into an interim role while MongoDB seeks a successor. A retained search is already underway, and the company plans to take the time needed to conduct it thoroughly.

A former BladeLogic founder, who became BMC Software’s president after BMC acquired BladeLogic, said he considered MongoDB’s CEO role after judging the database company ahead of rivals in developer mindshare and financial momentum. About six months after assessing competing database investments, he received a call from a search firm about the job. He later became a venture capitalist, led Datadog’s Series B round and remains its lead director.

He said MongoDB had strong engineering and product teams but a dysfunctional go-to-market organization and leadership culture. He believed newer-generation open-source technology was better than the earlier generation and that developers’ enthusiasm for MongoDB was an advantage. He also saw New York’s growing technology talent pool as making it easier to build a company there than when he started BladeLogic.

BladeLogic was founded in 2001, and its founder said the company raised its first financing round five days before 9/11.

He described the period as difficult for startups: customers and employees were wary of dot-com companies, capital was expensive, and BladeLogic had to build its own infrastructure.

A service-based model may be a simpler way to monetize open AI models than revenue sharing, according to an assessment based on MongoDB’s experience. MongoDB made the core product free for developers and charged for management tools, but drawing the line between free and paid features was difficult: giving away too much constrained monetization, while giving away too little could slow adoption.

Offering the product as a service made usage billable. Customers were described as paying $10 a month or “$1 million $100 a month”—an unclear figure—and could still download and run the product themselves. For AI models, concerns about intellectual-property rights may lead users to run models on-premises. Revenue sharing then raises questions about how much to pay; auditing users was described as infeasible, leaving the arrangement dependent on trust.

One proposed rule for startup governance is that management should lead when a company is performing well, while the board should take charge when it is not. Investors have little incentive to replace a founder, according to this view, and would consider doing so only if the company were doing very poorly. Founders are expected to meet their obligations; that may mean bringing in another CEO or a strong second-in-command. Constant fear of being removed by investors may indicate the founder chose the wrong partners.

Roelof Botha served on one CEO’s board for 11 years while that CEO was in the role and remains a board member; the CEO described Botha as challenging him at times and supporting him when needed. Separately, an observation about employee leverage used an illustrative case in which “like 50%” of a company’s enterprise value is tied to one person who is not a founder and may not even be on the management team. Such importance, the observation suggested, can give an employee economic leverage and greater influence over the company.

A personal assessment suggests Meta may warrant a valuation discount because it is unclear what major investment Mark Zuckerberg might pursue next. The scenario is conditional: if Zuckerberg gets AI working very well—with frontier models, leading consumer applications and an enterprise business—he might then bet another $500 billion on the next thing, whose success would be uncertain.

The assessment also holds that more stable management and some shareholder influence could probably lead Meta to trade higher, though the outcome is uncertain.

A CEO’s job is to set a clear strategy, assemble the right team and build a culture that helps the team succeed, according to an executive’s view. A clear direction makes it easier to set priorities and decide what the company will not pursue; as the business grows, the team may also need to change. The executive said business-building fundamentals matter more than a strong Rolodex, since professional connections come and go.

The executive also described culture as a product that can help attract and retain employees through connection to the mission, learning and meritocracy. CEOs should welcome trusted feedback that challenges their beliefs, reconsider strategy as markets change and make room for hard conversations, so the best idea—not the most senior person’s—wins.

Imprint says it is launching or relaunching Kroger’s co-branded credit card, with Kroger leaving a large bank for the fintech.

Imprint CEO Darr Murphy said the company is not a bank and works with partner banks; he described a $100 million investment to rebuild its Artech platform.

A co-branded credit card could give an AI agent a way to earn revenue from transactions, with possible rewards paid in tokens or other benefits and fees captured when users pay. The idea was discussed in the context of Instinct, a personal AI agent, as an alternative or complement to building partnerships such as one with Shopify; no card launch was announced.

The proposal also addresses a gap in referrals: ChatGPT can help users find local service providers that may have no referral program, leaving the service without a direct benefit. Card payments could link purchases to an AI account, while providers are already accustomed to transaction fees.

Imprint is investing in letting merchants accept bank accounts as payment. Its CEO argues that AI agents could make linking an account easier: users could give trusted agents their bank details and have them handle the connection. He sees this as a potential threat to card interchange, though no timeline or scale for such a shift was specified. He also cited “2% of American commerce” as going through “these,” calling it a huge amount of value; the excerpt does not make the reference fully clear.

The CEO said Stripe is promoting Link, citing an Uber checkout prompt, while Walmart’s earlier push to move customers to ACH was a long slog and Apple Pay gained ground. He views the shift as bearish for Mastercard and, with the caveat that it is a different business, American Express, but predicts it could benefit merchants and consumers over the longer term.

Outer Signal announced a $22 million Series A round today.

No investors, valuation or other terms were specified.

Outer Signal describes its platform as building detailed consumer profiles from checkout and transaction data so businesses can personalize marketing and identify who is buying. The company says resolving which person placed an order can take “hundreds and hundreds or thousands” of engineering hours. It presents the profiles as useful across email, paid media and customer acquisition, as well as retail strategy.

Outer Signal says it can help brands target audiences around stores and understand which customers drive sales. In one example, a creatine brand won a major retail placement after finding that a senior Walmart buyer had personally purchased its product. The broader market argument is that consumer brands increasingly plan for omnichannel sales from the outset and risk missing retail opportunities if they insist on being pure-play direct-to-consumer businesses. Outer Signal says it is attracting larger customers, including hotels, consumer apps and major retailers; an unnamed hotel company reportedly approached it after LeBron James’s stay went unnoticed until he reached the lobby because the booking was under his assistant’s name. The roughly one-year-old company says it acquires customers through word of mouth, ads, a Shopify app, agency partners, sales and content.

Instinct began sending product recommendations, and in one example the agent recognized that a user had a Targa from earlier work on car insurance but suggested a trickle charger and an indoor cover.

The recommendations were described as ill-suited to the car and the owner’s circumstances, underscoring that useful suggestions require both customer context and knowledge of the product.

Businesses use external market and customer signals not only for marketing but also for merchandising, product development and R&D. One large beauty brand found that about 10% of its customers were nurses, doctors or other medical practitioners. They bought one specific SKU dramatically more than the brand’s other customers, although the brand had not been targeting them. The insight could inform segmentation, marketing and product expansion for this group.

Real-time signals could also show which customer cohorts over- or under-index for a new online SKU. Mountain Valley was given as a hypothetical product-launch example.

Top Shelf’s founder says he started the fund in 2022.

It began investing in vice businesses, including alcohol and nicotine, and has since expanded into the broader consumer market and technologies that help businesses connect with their customers. He also identified Top Shelf as an investor in Outer Signal.

Gemini 4 Argon launched and is available to try.

Logan Kilpatrick posted about the release. Benchmark discussion was deferred until the following day.