Anthropic's IPO prospectus says second-quarter revenue rose 1,368% year over year, while its first-half operating loss narrowed to $693 million. The company has committed more than $600 billion to computing deals. The filing describes an independent trust that can appoint four of seven directors and founders holding a fixed 50.1% of voting power. Amazon and Google would receive no votes.
I can't write this story to the brief. The source text is a Bloomberg bot-check page, not the article. It has no reporting, no quotes, no named people and no dates. The only facts I have are in the headline: Physical Intelligence, a raise of about $1B, a valuation of about $20B, and that the company is in talks.
Anything that fills 120 to 180 words from that would have to be invented, such as investors, the company's product, previous rounds or timing. That would break your rule on unsourced figures and names.
Physical Intelligence is in talks to raise $1B at a valuation of about $20B. The talks are ongoing, and no deal has been confirmed.
If you paste the article text, or a different source such as a company statement or another outlet's report, I'll write the full piece in the house voice.
Dream, a startup building AI models for architecture, engineering and construction, has come out of stealth with $17 million in seed funding led by Andreessen Horowitz. SV Angel, XYZ and Zigg Capital also invested. Founders Andy Elder and Arjun Mathur were Stanford roommates who spent a decade at Palantir, where Elder led product design for a team of 800 and Mathur was a senior vice president.
The model is trained on more than 100 million square feet of building data from permit archives across the United States. The founders say it produces structurally precise designs that account for plumbing and HVAC. Dream aims to replace tools such as Autodesk's AutoCAD and Revit, and it is deployed across 10 active projects in Texas and California.
Motif, Snaptrude and Arcol are also applying AI to the sector, but Elder and Mathur say only Dream has a world model that instantly designs whole buildings. "I have not seen anybody take the technical approach," said Kimberly Tan, a partner at Andreessen Horowitz, which first invested over a year ago.
Reported by TechCrunch, via StrictlyVC, 2026-10-30
Anduril and Qualcomm announced on Tuesday that they will build BlueChip, a hardened laptop that will be open for preorder in 2027. Anduril says about 90% of its components will come from the US or allied countries, including a Snapdragon X Series processor built at TSMC's new facility in Arizona. Anduril has not finished contracting with all its component suppliers, and it has not announced pricing.
The first model is aimed at the military, government, first responders and law enforcement rather than consumers. It will be heavy and have battery-life constraints because it is built to withstand rain and dust. The selling point is a supply chain that customers can track and audit, which suits agencies required to buy US-made computers.
Palmer Luckey, Anduril's founder, says the display is one of several components made only in China, and that a technical roadmap exists to reach 100% domestic. He estimates a price about 20% above typical high-end PCs. For an ordinary buyer, he said, "I'd be better off buying a Dell laptop than an Anduril laptop." Rugged computers are already about a $5 billion market.
Reported by TechCrunch, via StrictlyVC, 2026-10-28
Google announced Gemini 4 Carbon on October 28 and gave access first to a subset of vetted cyber defenders, not to developers or consumers. The model follows Gemini 4 Argon by about two weeks. Google says it finds vulnerabilities, writes exploits and patches code better than Argon, and it has not given a date for wider release. Access runs through Fairwind, a programme Google says works with more than 650 partners.
The open question is transparency. Anthropic, which took the same route in April, publishes who is in Glasswing, while Fairwind shows partner logos without naming them in its text. A WorkOS analysis of the Argon rollout notes that vetting is a point-in-time check and that Google has not said what triggers a second review. Google reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, but nearly all the claims come from its own testing, and it has not cited an independent reproduction. Fairwind's page still lists Argon, so written terms for Carbon are not public.
Reported by Yahoo Tech, via Bay Area Times, 2026-10-29
More than half of Anthropic's customers spending at least $100,000 a year now use both the Claude platform and a first-party app such as Claude Code or Cowork, according to the company's S-1, filed with the Securities and Exchange Commission on Monday. At the end of last year the share was about 30%.
As of September 30, Anthropic counted more than 11,900 customers at that spending level, more than 1,900 spending at least $1 million, and about 250 spending $10 million or more annually. Net revenue retention was about 900% at the end of the third quarter. In that quarter, the top 5% of Claude Enterprise customers using Claude Code spent about 20 times as much per month as the average enterprise user without it.
The filing also lists risks. Amazon, Google and Microsoft increasingly see Anthropic as a competitor, which it says could complicate access to the compute it needs from them. It also warns that losing one or more of its largest customers could hurt its business. Revenue for the first half of 2026 was $16.3 billion, against $1.2 billion a year earlier.
Reported by TechCrunch, via StrictlyVC, 2026-10-27
Some 230 organizers and political workers at SEIU-UHW, the union that put Proposition 40 on the California ballot, have gone on strike against their own employer days before polls close on Tuesday. The walkout has stopped all campaign work for the measure, which would levy a one-time 5% tax on an estimated 200 billionaires who lived in California on 1 January 2026. Pickets are under way at the union's offices in Oakland, Los Angeles, Sacramento and San Diego.
The staff, represented by SWU Local 2599, say they have gone two weeks without a labour agreement after nine bargaining sessions. They want 5% raises in each year of a four-year contract. Management has so far agreed to a 3% floor. Wilson Saiki, a field organiser, said: "This is the most important week for our election and SEIU-UHW forced us to be out here on strike." Spokesperson Renée Saldaña said the union looks forward to resuming bargaining shortly, and that work on Prop 40 continues through the Billionaire Tax Now coalition.
The Commodity Futures Trading Commission is again examining whether Polymarket broke a 2022 settlement that required it to keep U.S. users off its offshore prediction market. The inquiry covers whether the company's geolocation controls and its U.S. advertising were sufficient. The CFTC and the Justice Department closed similar investigations last year without bringing charges.
Of the 123 local measures San Franciscans have voted on since November 2016, about 70 percent would have been unaffected by Proposition D, according to a Mission Local analysis. Only 18 percent of local initiatives in that period reached the ballot by signature-gathering, and more than half were placed there by a majority of supervisors.
The measure would raise the signature threshold from 2 percent to 8 percent of registered voters, roughly 10,600 to 42,500 signatures. It would raise the number of supervisors needed to place an initiative from four to six, and end the mayor's unilateral power to put measures on the ballot. The city's Budget and Legislative Analyst's Office found that qualifying an initiative would cost about $1.5 million, against some $320,000 now.
The "Yes on D, E & F" committee has received nearly $14 million. Opponents, including several labor groups, have raised $206,000. Mayor Daniel Lurie says the measure will give voters a stronger check on power. "Don't deny citizens access to the ballot," said Dale Carlson, a longtime political consultant. Voters decide on 3 November.
California's SB 53 took effect on 1 January 2026. It applies to developers that have trained or begun training a model using at least 10^26 FLOPs, with a stricter tier for large developers that also had gross annual revenue above $500m in the previous calendar year. A METR note dated 29 January 2026, by Miles Kodama, Michael Chen and Garrett Xu, summarises the law alongside New York's RAISE Act, Illinois' SB 315 and the EU AI Act.
Large developers must publish a frontier AI framework on their websites, and its commitments are legally binding. A developer that fails to follow its own framework can be fined up to one million dollars per violation. Critical safety incidents go to the California Office of Emergency Services within 15 days, or within 24 hours where there is imminent risk of death or serious injury. Large developers also submit quarterly summaries of catastrophic risk assessments from internal AI use. The note does not describe any requirement to report the share of compute spent on safety. RAISE and SB 315 both take effect on 1 January 2027 and require incident reports within 72 hours.
Reported by metr.org, via a16z AI Policy Brief, 2026-11-02
On Sept. 5, cryptocurrency billionaire Chris Larsen gave $500,000 to a political action committee backing Theo Ellington in the District 10 supervisor race. On Sept. 29 he added $40,000, which he told Mission Local in a text message covered the expenses of the team managing the PAC. Ellington is the frontrunner. He has the endorsement of Mayor Daniel Lurie and nearly $488,000 in direct donations and public financing.
His opponents, Dion-Jay "DJ" Brookter and J.R. Eppler, have launched a joint "anti-billionaire" campaign and are asking supporters to rank each other second. Ellington says he does not control the PAC and that the fallout has made the contribution more trouble than it is worth. Aaron Peskin, a former supervisor who backs Brookter, said of the donation: "They're doing this guy no favors."
Todd David, a moderate political consultant who donated to Ellington, said the ranked-choice strategy has rarely worked. Many voters remain undecided. In 2018, turnout in the district was 58%, against 78% citywide.
Binance suspended withdrawals on multiple networks on 29 October after users reported funds being drained from the exchange's wallets. Crypto watchers say at least $2 billion has left its hot wallets, which are connected to the internet. Some estimates put the loss above $5.7 billion, but the figures are unconfirmed and the theft may be continuing.
Binance spokesperson Ross O'Leary said withdrawals were suspended "as a precautionary measure". The company says it is investigating and that user funds are safe. Founder Changpeng Zhao, now its largest shareholder, said on X that the company had likely been hacked and that he expects users to be fully covered, without explaining why.
Circle froze about $800 million in Tether passing through its networks, and Tether froze about $1.56 billion tied to the hack. It is unclear who is responsible, and Binance did not say whether the hackers had contacted it. Bybit lost a record $1.4 billion in a single breach in 2025, which was blamed on North Korean hackers.
Reported by TechCrunch, via StrictlyVC, 2026-10-30
OpenAI has released GPT-6.1 Terra, the first Terra-class model since GPT-5.6, according to Artificial Analysis, which published its results on 26 October 2026. Pricing is $1 per million input tokens and $5 per million output tokens, against $2 and $12 for GPT-5.6 Terra. At max effort the model gains 6 points on the Artificial Analysis Intelligence Index, at $0.40 per task compared with $1.40.
Not every setting sits on the cost frontier. Low and medium effort are Pareto-optimal on the Intelligence Index, while the other effort levels fall behind it. At max effort the model scores 4 points below GPT-6.1 Sol and 10 above GPT-6 Luna. On the Coding Agent Index v2.0 it scores 48, against 54 for Sol and 31 for Luna.
In AA-Omniscience, the model guesses 38% of the time when it does not know the answer, down from 88% for GPT-5.6 Terra. Accuracy rose 2 points. Terminal-Bench 4.0 improved by about 20 points, while most other evaluations moved less. GDP.pdf gained 5 points.
Hackers stole at least $6.3 billion in crypto from wallets at Binance, the world's largest exchange, according to security researchers. About half of that was frozen as the funds moved across blockchains. TRM Labs and Elliptic both attribute the theft to North Korean hackers, and TRM says it took about 20 minutes.
Binance said in a post on X that attackers reached a service container through a compromised engineer's laptop and submitted unauthorised withdrawals. It has seen no evidence of a private key leak, has suspended all deposits and withdrawals, and says user funds are safe, though it has not yet provided proof. Spokesperson Andie Cooper said about $3 billion has been frozen by industry partners and that Binance has "begun recovering affected assets for users".
It is unclear how the laptop was compromised, or whether Binance commissioned an outside security audit before the attack. The incident is not confirmed as resolved, and the stolen total may change. Researchers also blame the group, known as TraderTraitor, for the $1.4 billion Bybit theft in 2025.
Reported by TechCrunch, via StrictlyVC, 2026-10-31
Rockwell offers free benefits concierge to make money on the network beneath it
Rockwell is building an AI-native benefits department for mid-market employers. The company, which is part of the Daybreak portfolio, offers a free care concierge that helps employees navigate their plans. It also aggregates demand toward providers, pharmacy programs and care-management services. The summary gives no figures, customer names or launch details.
Benjamin Todd of 80,000 Hours updated his article on AGI timelines on 23 October 2026. He puts the chance of artificial general intelligence arriving by 2035 at higher than 50 per cent. He thinks the chief executives of the leading AI companies are too optimistic, but says there is a serious argument for AGI by 2030.
The article says the length of coding, machine learning engineering and cybersecurity tasks that models can complete doubled every seven months before 2024 and now doubles every four. In October, AI solved hundreds of important mathematics problems, many of them unsolved for decades. Demis Hassabis of Google DeepMind said in June 2026: "We're only a few years out now."
Todd argues that investment in computing power and algorithmic research cannot keep growing at current rates much beyond 2030. Either AI reaches the point where it can sustain its own progress, or progress slows. He also sets out a case against, in which models never handle ill-defined, unverifiable work over long time horizons and stay a tool. Yann LeCun, who doubted in 2022 that such models could learn basic physical reasoning, now expects AGI in several years.
Axiom Math has released what it describes as the first complete computer-checked proof of the Classification of Finite Simple Groups. Over 24 days, its engineers and mathematicians directed swarms of dozens of AI agents that produced approximately 43.1 million lines of code in the Lean proof assistant. The original proof runs to roughly 15,000 pages across more than 500 articles by about 100 mathematicians.
Translating the work into Lean exposed problems in the published arguments. Most were minor, such as typographical errors, errors in tables or dropped hypotheses, but 34 noteworthy deviations had to be resolved. The most substantial concerned the Unbalanced Group Theorem. The work also produced a new result in group theory. Axiom says it consulted several expert group theorists and will make the lengthy errata list available.
The team formalised the first generation proof, because the second generation proof remains incomplete. Michael Aschbacher, who co-published the 2004 treatment of the quasithin case, wrote in 2005 that "it is impossible to write out a very long and complicated argument without error."
OpenAI's GPT-6.1 Terra tops Grok 4.7 and MiMo-V2.6-Pro on AA index... and? I don't want to knock the engineers achievements but in the real world what does that translate to?
More importantly who is buying it? Is the race to beat the benchmarks actually losing sight of the commercial needs of the average client? Why pay more for a top performing model if a middle tier offering provides the outcome you need for half the price?
Frontier AI companies need to realise at some point they need to make a profit. Right now retail investors are the exit plan! Time to start looking at the product not the engine?
