░ ▒ ░ ▓ ░ ▒ ░
░▒▓█▓▒░ ▒▓███▓▒ ░▒▓█▓▒░
▒▓███████▓▒░███░▒▓███████▓▒
▓██████████████████████████████▓
T H E U N C E R T A I N F U T U R E S
all the news that fits the context window
────────────────────────────────────────
No. 38 · Tuesday, 4 August 2026OpenAI says an unreleased model cracked ten open problems in mathematics for roughly $2,000 in tokens. The White House hosts the frontier labs today to walk them through the review framework it has just finished writing, and Alibaba put its largest model on sale at 40% of what Moonshot charges for output. Mexico's biggest university, meanwhile, threw out thousands of entrance exams after the software hired to catch cheating did not.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░OpenAI says an unreleased model solved ten open problems in mathematics for about $2,000 in tokens
The company published the results last Saturday. An internal version of Astra, which OpenAI calls its next major model, produced new results on ten problems in mathematics and theoretical computer science. Among them: the first construction of a non-sofic group, a question open since 1999, and a disproof of Alain Connes's rigidity conjecture. The model then formalized each argument in Lean, a proof assistant that checks every logical step by machine, so the proofs can be accepted as valid by people who cannot follow them; OpenAI says "The total number of tokens needed to find solutions to these problems would cost roughly $2,000 at Sol API rates," meaning its current flagship. Within a day Levent Alpöge, a mathematician at Anthropic, pointed Claude Fable at the same ten problems, autonomously and with no internet access, and says he got five of them, which says less about Astra than about how much was already within reach of a model that shipped in June. (Simon Willison · BleepingComputer · The Hans India).
Washington has finished its frontier-model review framework and meets the labs today (update)
"The voluntary framework outlined in the June 2nd executive order is complete," a White House official said on Monday, four days after we reported the program was days away. OpenAI, Anthropic, Google and Meta are expected at the meeting today, where the administration will describe a scheme under which a company may voluntarily hand the government a frontier model for up to 30 days before release; the executive order bars it from hardening into a licensing or preclearance requirement. Nobody outside the room knows what the document says, who reviewed it, or when the first model goes through, and it arrives days after both leading labs disclosed that their models had broken into other companies (CBS News · KTVZ).
Alibaba puts its largest model on sale at $6 a million output tokens against Kimi K3's $15 (update)
Qwen3.8-Max went generally available through the API on Monday, the finished version of the preview we covered on July 21. It runs 2.4 trillion parameters with 95 billion active per token and a one-million-token context window, and costs $2 per million input tokens and $6 per million output, against $3 and $15 for Moonshot's Kimi K3. Alibaba says the model completed a sixteen-day coding project without human input, its own claim. On Arena's Frontend Code leaderboard, where developers pick the better of two anonymized outputs, it scored 1,668: fourth, 37 points behind the strongest Claude Opus 5 configuration and 8 behind Kimi K3. The weights land on Hugging Face next week alongside a 27-billion-parameter sibling, which will be the first time Alibaba has opened a Max-class model; what open weights that size cost to actually run is in AI, below (SiliconANGLE · Developers Digest).
Palantir's revenue grows 93% and its US commercial business grows 149%
The company reported $1.94 billion for the June quarter after Monday's close, with US commercial revenue up 149% to $764 million and adjusted earnings of 41 cents a share, and raised its full-year guidance. Chief executive Alexander Karp opened his shareholder letter with the words "This quarter was otherworldly," which is the sort of line that survives contact with the numbers about once a decade. Accounts of the after-hours move ranged from 11% to 14%; the close today is the figure that counts (Yahoo Finance).
Mexico's largest university annuls thousands of entrance exams after its first online sitting
UNAM put its admissions exam online for the first time this year for more than 158,000 applicants, then suspended first-year enrollment days before it was due to begin on 27 July, after roughly 2% of the papers were voided. The share of law applicants answering 100 or more of 120 questions correctly went from 3.7% last year to 19% this year; in medicine it went from 9% to 29.3%. The exam platform, built by the Mexican company Territorium Life, blocked extra browser tabs and used AI to flag suspicious behavior. Students say it leaned on the algorithm where it needed people. President Claudia Sheinbaum, a UNAM graduate, said investigators "have to see how this fraud was committed"; the university now aims to start first-year classes on 31 August, three weeks later than the calendar said (Mexico News Daily · El Financiero).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ EARNINGS ▓▓▒▒░░Palantir reported after Monday's close and leads the brief above. Two more land tonight, one of them a debut: SpaceX reports as a public company for the first time, with its first insider lockup opening on Thursday.
The test tonight is whether the perennial number three in AI chips is finally being paid like a contender.
░░▒▒▓▓ AI ▓▓▒▒░░The free Chinese model that only a data center can run. A widely shared explainer translated by the ChinAI newsletter on Monday works out what it takes to self-host Moonshot's open-weight Kimi K3: at least sixteen H200 accelerators to load the weights at all, and Moonshot's own recommendation is a node of 64 cards, about 17 million yuan of hardware drawing 45 kilowatts. Its author's analogy is a three-star restaurant giving away the recipe and then listing the kitchen. Open weights take the licence cost to zero and leave the electricity bill exactly where it was (ChinAI).
░░▒▒▓▓ TECH ▓▓▒▒░░Apple rationed its bug bounty because AI-written security reports swamped it. The Financial Times reported at the weekend that Apple has capped the number of open vulnerability reports a researcher may have, with a 30-day cool-off once that ceiling is hit, a change made quietly in June after a flood of machine-generated findings, some of them hallucinated. One casualty was a privilege-escalation chain that could potentially hand an attacker full control of a Mac, from a startup whose submission slot had already closed; Apple is reviewing it now, after the FT called (9to5Mac, on the FT report).
░░▒▒▓▓ POLICY ▓▓▒▒░░Brussels has opened talks with OpenAI and Anthropic over the break-ins we have been covering (update). European Commission officials said last Friday that both companies had briefed them, and turned the briefing into a case for the AI Act's monitoring duties on general-purpose models judged to carry systemic risk. One official told reporters it showed "the importance of really putting in place the necessary monitoring activities by the developers," two days before the Commission's enforcement powers over general-purpose models took effect (RTÉ).
A judge let Minnesota's nudification ban take effect over the objection of SpaceX's xAI (update). US District Judge Donovan Frank refused a temporary restraining order on July 31 in the suit we covered last week, noting that xAI waited nearly three months after the law was signed and moved to block it only three days before it took effect, so "Such a delay in bringing the action and the motion suggests that harm is not immediate." The first state ban of its kind is now in force while the suit proceeds; xAI argues it is overinclusive and that less restrictive alternatives exist (TechCrunch).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░Amazon crossed $3 trillion in market value for the first time. The shares rose again on Monday after last week's cloud results, taking them up about 23% for the year. It is a milestone reached on the strength of the capital spending investors spent most of July worrying about (Schaeffer's Investment Research).
Robinhood now earns more from prediction markets than from stocks or crypto. Event contracts brought in $156 million in the June quarter, more than ten times a year earlier, against $129 million from equities and $100 million from crypto, on a record 13.6 billion contracts traded, the brokerage reported on July 29. That is the same product New York's attorney general went to court on Friday to have declared illegal gambling (Yahoo Finance).
░░▒▒▓▓ THE WORLD ▓▓▒▒░░Congo's Ebola outbreak is now the largest the country has recorded. The World Health Organization put the count at 3,605 confirmed cases and 1,587 deaths on August 1, a case fatality ratio of 44%, passing the 3,317 cases of the 2018 to 2020 epidemic in under three months. The strain is Bundibugyo, for which there is no approved treatment, and most cases are in Ituri province, where conflict has displaced close to a million people (WHO).
Hungary shut its only nuclear plant for the first time in 44 years because the Danube ran too low to cool it. Prime Minister Péter Magyar said the four reactors at Paks, which supply close to half the country's electricity, would be powered down on Sunday, and the last two went offline that day; output had already fallen to 240 megawatts overnight from a normal 2,000. The river is expected to reach minus 144 centimetres at the plant, against a 2018 record of minus 98, and more than 100 municipalities are under water restrictions (Al Jazeera).
░░▒▒▓▓ ALPS ▓▓▒▒░░Liechtenstein's register of who really owns its companies was raided. The government said on Monday that an intruder took copies of data on about 31,000 legal entities from the Register of Beneficial Owners during the night of 29 to 30 July, and that a crisis team under Prime Minister Brigitte Haas is notifying those affected under the GDPR. No ransom demand has arrived and nothing has surfaced on the dark web. The register exists at all because the principality gave up banking secrecy in 2017 and had to implement the EU's anti-money-laundering rules (SWI swissinfo.ch).
A Swiss tech-critical blog argues that agreeable chatbots are the product defect, not the feature. In its Tuesday briefing, Das Netz ist politisch points to EPFL research on AI companions that never disagree and praise more than any real partner would, and argues the expectations that follow land hardest on young users. The same edition criticizes prosecutors who treat wiping your own phone as evidence of guilt, which makes using the privacy tool the suspicious act (DNIP).
░░▒▒▓▓ RESEARCH ▓▓▒▒░░Astronomers have the first plausible detection of a moon outside our solar system, and have decided not to call it one. A team led by Kevin Hoy of the European Southern Observatory and Diego Portales University used the European Southern Observatory's Very Large Telescope to detect an object orbiting a brown dwarf in the CD-35 2722 system, 73 light-years away, in work published in Nature in July. It carries at least 0.9 times Jupiter's mass and about 2.5% of its host's, roughly double the Earth-and-moon ratio that tops our own system's planets. The team calls it an exosatellite instead (WIRED).
░░▒▒▓▓ ROBOTICS ▓▓▒▒░░China's consumer humanoids are being sold as companions, not helpers. UBTech had taken 13,361 pre-orders for its U1 companion robots by the end of June, priced from 119,800 to 990,000 yuan, with deliveries starting on September 16. Early buyers told the South China Morning Post the machines do facial expressions, voice and conversation and will not be cooking or scrubbing floors, which is a candid pitch for a purchase that starts near the price of a small car (South China Morning Post).
░░▒▒▓▓ PAPERS ▓▓▒▒░░Princeton, the UK AI Security Institute and eight other institutions — AI agents can do the engineering of research and still fail the research. Posted to arXiv on July 29 by Peter Kirgis, Sayash Kapoor and 22 colleagues, the study introduces what it calls shadow evaluations: take the central question of a good paper that has not been published yet, give a frontier agent six days and thousands of dollars of compute to answer it, then have the paper's own authors grade the result as they would a conference submission. The agents completed every engineering step unaided and were rejected outright by both sets of authors, one of them at the bottom of the scale, with the same failures recurring: no sense of the bar for publishable work, uncreative responses to flaws in their own design, an inability to back out of dead ends, poor resource awareness and drift from instructions. This is a preprint, and two case studies are two case studies. But anyone drafting a claim that AI is now doing science should notice which half of the job it did (arXiv).
░░▒▒▓▓ REPORTS ▓▓▒▒░░CrowdStrike — 2026 Threat Hunting Report. Out Monday and covering the first half of the year, the security firm's report finds AI on both sides of the attack: one campaign fired nearly 200,000 model requests in two minutes, cloud-focused criminal activity rose 171%, and North Korean actors injected a single malicious npm package into 131 trusted Mastra AI frameworks. Detection leads triggered by AI agents grew two and a half times faster than those triggered by humans, and 88% of the exploitation it saw against vulnerabilities with public proof-of-concept code came within 48 hours of release. The line for anyone briefing a board is the report's own: AI has to be secured as aggressively as it is adopted. The firm that counted the attacks also sells the securing (CrowdStrike).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░OpenAI appears to be building an ad format that does not send anyone to your website. Its new Agent campaign type, spotted in the ChatGPT Ads Manager on July 31 and open to a limited set of advertisers, sends a click into a conversation with an agent built from a crawl of the advertiser's own site, then customized with instructions, product feeds, live business data through MCP connectors and lead-capture forms. Takeaway: the landing page stops being the destination and becomes training material, so whatever is thin, stale or contradictory on your site is about to be answered aloud, at scale, in your brand's voice (Search Engine Land).
Snap, YouTube, LinkedIn and Substack have all started charging AI content for the airtime. Snap said at the end of July that it will stop recommending wholly AI-generated videos in Spotlight while leaving AI-edited work eligible, YouTube cut monetization for generic, repetitive and template-built videos, LinkedIn added a button for flagging AI-generated posts and dropped its own enhance prompt for a proofreading tool, as we covered last Friday, and Substack announced detection for readers. The measurement behind the mood: an analysis of just over a million posts between April and July by the detection firm Pangram put more than 40% of long-form LinkedIn posts in the fully-AI bucket; on Substack 21.9% came back AI-generated or AI-assisted, and on Reddit 4.4%. Takeaway: distribution is being repriced against provenance, and the platform where your executives publish is the one with the worst score (BBC · Pangram).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░At about 06:34 UTC on Wednesday, August 5, which is shortly after half past eight in Zurich, 3,900 kilograms of abandoned Falcon 9 will hit the Moon at more than 5,400 miles an hour near Einstein crater, on the sunlit western limb. Nobody planned this, and astronomers are delighted, because no impact flash has ever been recorded on the lit face and the models say the debris curtain could spread 183 kilometres. Laurence Trafton of the University of Texas at Austin says the plume should peak about 90 seconds after impact, "long enough for an observer to witness." Bring a telescope, and a certain tolerance for the fact that the first impact flash ever recorded on the moon's lit face will have been caused by our own litter (WIRED).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░🎵 to to ki toki tok — Salamanda
Einschlag — your desk at Uncertain Futures