We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
The Riemann proof has passed its formal checks
Last brief, Levent Alpöge's Claude-assisted proof of the Riemann hypothesis was still waiting on outside checks. Those checks have now come back. The Lean community posted a statement confirming four things :
- The Riemann hypothesis statement used is the one already in Mathlib, which experts had vetted.
- Dozens of people compiled the library and got the same proof export.
- Two independent tools confirmed that the theorem Alpöge proved is exactly Mathlib's Riemann hypothesis.
- The exported proof passed Lean's kernel and several independent checkers. Two of those checkers come with formal guarantees against accepting a false proof.
Exponential View notes that Alpöge's proof runs to 17.4M lines of Lean. A separate all-human proof by Alessandro Masullo is being translated into Lean so it can be checked faster .
For investors, the takeaway is that machine checking, not human refereeing, is becoming how frontier results get validated, even for human-written proofs. The proof is certified. Whether mathematicians can actually learn anything from a proof this long remains open.
AI M&A: $4.5B for a team with no disclosed revenue
Palo Alto Networks is buying Applied Compute, which does post-training and inference for open-weight models, for $4.5B in stock. The release gives no revenue or customer figures . CEO Yash Patil becomes Co-Chief AI Officer, and the team will train Palo Alto's own models .
Nikesh Arora's reasoning is that enterprises will run a mix of open and closed models, and every one of them will need securing. Palo Alto first saw the platform as a prospective customer .
The buyer's own numbers are worth a look. Acquisitions drove most of its 39% Q1 growth, and growth falls to about 18% once the CyberArk deal is in both comparison periods . GAAP operating income was only $96M on $3.44B of revenue, because deal costs and stock compensation are piling up .
Perplexity passes $1B in run-rate revenue
Aravind Srinivas says Perplexity Computer, launched in February, drove 4.7x year-over-year revenue growth . The company crossed $1B ARR after targeting $650M by year-end. He describes a 2028 IPO as the "mental model," and says it could file sooner . He cites 56% gross margins and no large compute commitments . Routing simple answers to its own post-trained model saves about $100M a year .
The catch is model dependence. Perplexity pitches itself as model-neutral, yet Opus 5.5 handles 60–70% of agent tokens. Srinivas concedes Perplexity lags on how efficiently agents complete tasks, and says that is why it is building a post-training research lab .
Compute stays scarce, and power is the bottleneck
Kush of Ornn, a market for AI compute, told 20VC:
- Supply: buying capacity from CoreWeave today can mean a Q4 2027 or Q1 2028 delivery quote. Building your own takes about six months .
- Prices: renewal options at around $2.50 per GPU-hour are no longer free. He wouldn't be surprised by $20–30 per GPU-hour in three years . The risk to that view is weaker enterprise demand pushing GPU "rents" down .
- Where he'd invest: he calls "a round done in 2 hours" FOMO-driven . He says the vertical-AI categories now led by companies like Harvey and Rogo are already taken . He would back unglamorous physical infrastructure, such as a startup building better transformers .
Exponential View adds that more than 2,000 GW of US generation and storage is waiting to connect to the grid, which is more than all plants running today . With an AI chip renting for about $3 an hour, a month spent waiting for power costs more than the power itself. On-site solar plus batteries is the cheapest data-center power option in half the US .
Data leaks among inference providers
OpenRouter off-boarded three inference providers after security researchers reported compromised data or data policies not being upheld. It also removed 125 endpoints from Zero Data Retention (ZDR) eligibility, some as a precaution. OpenRouter says it found no breach of its own systems and credited a Dragos report .
Suhail is pitching Particle on that platform with explicit promises: it respects ZDR and will never route traffic to China . A researcher Suhail credits with uncovering many of the breaches confirmed there were "no leaks through particle from july to november" . Data handling is becoming a selling point inference providers compete on.
Policy on Chinese models and compute
Clem Delangue praised a new US House report for recognizing that the right answer to Chinese open-source AI is to build more American open source, not to ban it .
A WSJ headline says Southeast Asian data centers exploited a sanctions gap to give Chinese firms compute and cover for alleged chip smuggling, and frames Jensen Huang as the builder of that route . One commenter disputes the framing, saying the route was legal until it was banned on November 19 .
Fundraising timing and revenue metrics
Elizabeth Yin's advice for the holidays:
- If you are already raising, close quickly and take what you need to reach the new year .
- If you are in trouble and haven't started, run a small process aimed at fast-moving emerging managers .
- Only companies with standout growth tend to close in this window .
Separately, an investor she quoted says many vertical-AI startups exaggerate revenue, and claims to have seen 15 definitions of "CARR" . Ask founders exactly how they define it during diligence.
Agents: more activity than shipped output
- NBER study: a paper covering 500k+ GitHub developers reportedly finds autonomous coding agents raised commits by about 240% but releases by only about 30% .
- Salesforce ROAM: its release-review agent cut median turnaround from about 24 hours to 12 seconds across 18 products. Code decides whether a patch passes, the model only explains the result, a human gives final approval, and about 8% of requests go to a person .
- LangChain: it is promoting a "knowledge factory," a curated company knowledge base maintained for its agents .
Elsewhere
- AMI Labs: Yann LeCun's Paris company builds world-model AI rather than text predictors. It raised just over $1B in March . He argues extinction warnings help the biggest labs win rules that shut down open research .
- Pharma: GSK, BMS, Novartis, Lilly, Novo, AstraZeneca and Sanofi say AI has cut by as much as half the time it takes to produce new molecules and move them into the clinic. They now use it to design and rank molecules and to flag safety risks .
- OR cofounder and CEO Kush describes building a physical GPU-compute market to generate standardized reference indices, with futures and options on top. Because neoclouds would not share transaction data, OR began using third-party capital to control and operate chips, including through sale-leaseback structures. He sees compute-market infrastructure as a potentially trillion-dollar opportunity.
- Kush says GPU supply is tight: CoreWeave capacity was being quoted as far out as Q4 2027/Q1 2028, while a self-build takes about 20–22 weeks to obtain chips plus 2–3 weeks to install. He says GPU prices have been rising with AI demand, but identifies enterprise ROI failure as a downside: overbuying could leave tenants unable to pay and chips worth only their material value rather than their value as usable compute. Inference tokens have paying users, unlike more speculative training, but he flags neoclouds lending to startups without enough tenant diligence as exposed if those tenants fail.
- He expects older A100/H100 GPUs to retain use as rented capacity and move into lower-scale inference. His market thesis is one horizontal base provider plus many specialized inference companies—for example, differentiated by speed, cost, or workload—with provider count potentially shrinking as the market verticalizes.
- Kush views the “round done in 2 hours” dynamic in AI as FOMO rather than careful diligence, and says customer support and other “easy” application categories are crowded, citing Rogo and Harvey. His underheated area is the physical AI and energy stack, especially data centers, electricity infrastructure, and transformers; he says he would invest in a startup building better transformers.
- OR has about 28–29 people; Kush says scaling toward 100 requires communication and competitiveness, and that hiring should assess cross-functional fit and ability to handle uncertainty—not just technical skill, given AI tools and the company’s non-frontier-research focus.
- Perplexity says its cloud agent Perplexity Computer, launched in February, drove 4.7× year-over-year revenue growth. Business usage is the main source of revenue growth so far, concentrated in small and midsize firms; leading demand areas include finance, legal, corporate strategy and development, web-app building, and e-commerce, with large-enterprise expansion a next focus.
- Perplexity’s stated differentiation is search-backed accuracy and model-neutral orchestration; it claims a trillion-page index, 56% gross margins, and relatively low burn without huge compute commitments. It says routing nearly all standard-answer requests to its own post-trained model saves $100M annually. A caveat to its model-neutral positioning: Opus accounts for 60–70% of agent tokens, and the company says it is still behind on agent-trajectory efficiency as it invests in post-training, RL, and a Frontier Research Lab.
- Srinivas said Perplexity had crossed $1B after a $650M year-end target; the discussion framed the figures as run-rate ARR. He described 2028 as the company’s IPO mental model, while saying it might file sooner if progress sustains.
- LeCun argues that text prediction alone does not give AI a model of the physical world. After leaving Meta at the end of 2025, his Paris-based AMI Labs pursued AI that learns physical-world models rather than predicting text and raised just over $1 billion in March.
- LeCun identifies concentration of power, including AI-mediated information diets, as a major concern; he argues that extinction warnings can help the largest labs win rules that restrict open research and entrench their lead.
- AI power bottleneck: More than 2,000 GW of new US power plants and batteries are waiting for grid connection—more than the capacity of all power plants currently running in the country—and about 85% of the queue is solar, wind, and storage; the author says the queue is pushing AI off-grid. Solar modules are getting 24% cheaper with each doubling of capacity, and the briefing says a month waiting for power costs more than the power itself when an AI chip rents for about $3/hour; an Investment Brief analysis found on-site solar and batteries the cheapest data-center power option in half of the US.
- Anthropic investment framing: The briefing says Wall Street got “a piece” of Anthropic but does not specify what was bought; it frames the investment question as whether Anthropic can capture value across the intangible economy, with scenarios ranging from licensing models that run on devices to becoming an AI utility.
- Human-made content as differentiation: Rockstar’s Grand Theft Auto VI was reportedly made without generative AI by 5,000 staff, including 30,000 hand-crafted unique characters, and generated $3.79 billion in sales by the end of its first day, including preorders. The author sees a possible mass-market “certified human” tier emerging amid synthetic-content saturation, rather than human-made work remaining only a luxury niche.
- AI-assisted mathematics, with verification caveats: Anthropic mathematician Levent Alpöge says he and Claude may have proved the Riemann hypothesis; he says the Lean proof runs to 17.4 million lines. Separately, a proof by Alessandro Masullo, described as all-human, was being checked by a third party, with a team translating it into Lean; the author says he had not read either proof.
Paul Graham says tech insiders are unusually unanimous that a board made a major mistake regarding Jensen Huang, while the board disagrees; the post does not identify the board or decision, making this a broad sentiment signal rather than a specific governance development.
- Stebbings describes 23-year-old Kush as having built a viral sports-betting platform in his MIT dorm before starting OrnnExchange, a financial market for AI compute; he calls the company “insanely profitable.”
- Kush says compute is severely undersupplied: a buyer seeking capacity from CoreWeave today may be quoted Q4 2027 or Q1 2028, while building independently takes about six months, including 20–22 weeks for OEM chips and 2–3 weeks for installation. He says the market no longer gives away options to renew GPU-hour contracts at prices such as $2.50, and he would not be surprised by prices of $20–$30 per GPU hour in three years.
- Kush’s infrastructure-investing thesis compares data centers to rental properties: GPU landlords collect rent from inference providers and model companies while they remain tenants, but weaker enterprise demand could lower GPU prices, as rents fall in a weak economy.
- From Ornn’s experience, Kush says Cognition’s product was too expensive, views Factory and Cognition as competitors, and says Factory is much cheaper to use; he also rates Factory’s underlying harness above Devin’s.
- Kush argues boutique investors may be more founder-aligned than mega-fund investors because their success is more directly tied to portfolio companies and their firms are working to build their reputations.
a16z’s Moses Sternstein argues Gen Z is not a uniform cohort: in a comparison using 13 policy questions, 8.5% reportedly found someone with identical answers among 99 same-country Gen Z peers, and the likelihood of a match changed little when compared with other generations. Sternstein says cross-generational matches were slightly more likely with Gen X and Boomers than with Gen Z peers, attributing fragmentation to differences in internet consumption; he says this makes reaching young audiences difficult, a caution against treating them as one consumer segment.
Garry Tan says he has been using Capy for about half of the work with his software factory, calling it his best experience so far across desktop, mobile, and cloud. Capy is promoting an iOS app and inviting people to comment for TestFlight access.
Perplexity CEO Aravind Srinivas praised the “orchestration” behind a user-requested 3D history of technology spanning 3.3 million years; the site contains 4,800+ source-linked records, lets users compare past predictions with outcomes, and uses custom JavaScript rendering with logarithmic spacing.
A YC startup’s GBrain is described as automating team workflows: it files sales calls into team memory by 7 p.m., filters both founders’ inboxes three times daily, turns buyer feature requests into product pages linked to the originating calls, and reports nightly on changes in staging versus production; it produced 850+ pages in three weeks.
Garry Tan speculates that, with enough tokens, a hobbyist could use “vibe coding” to make a PC port of GTA VI that removes its disliked design flaws within a month or two; this is a forecast, not a reported project or release.
Martin Casado says his home assistant now speaks through an avatar on an iPad; getting it working required running a Grok bot and avatar server together on a Mac mini for audio forking, then routing audio between the Mac mini and iPad. He describes the home AI as still coming along.
Sriram Krishnan called GTA VI an apparent technical leap in graphics and game development and asked for technical analyses. The accompanying assessment by Tim Soret credits the achievement to complex scenes with many simulated NPCs and physical objects, hair simulation, volumetric lighting, global illumination, ray-traced reflections, and seamless data streaming at native 1440p.
A teammate built Getsumen Express, a browser-playable Moon-and-Mars exploration game, on Fable 5.2; its workflow combined Three.js with WebGPU/WebGL2, Blender and Python-built stations and train, Tripo/Meshy image-to-3D assets, and procedurally generated terrain, landmarks, and audio. The reported Fable 5.2 xhigh cost was about 73,000 credits (~$730).
Particle is on OpenRouter; Suhail said it will respect ZDR, never reroute traffic to China or send data elsewhere, and continuously monitor security. In a follow-up, he said every GPU Particle has is in the U.S. He cited more than a decade at Mixpanel without major issues and said he maintained a high bar for security and privacy there.
After the November 13 FAA ground stop, car and hotel bookings on Navan rose to roughly twice normal hourly volume as travelers sought alternatives, a signal of demand for travel-booking services during disruptions.
Jason Calacanis said he was buying more $NVDA that week, grouping it with chickens and ammunition.
Elizabeth Yin mocked aggressive run-rate extrapolation with the example, “I found a dollar in just 1s on the street, so my runrate is $7.2m.” The investor she quoted alleged that many vertical-AI startups exaggerate revenue to justify fundraising and claimed to have seen 15 definitions of “CARR”—a metric-diligence warning, not verified evidence about the market as a whole.
Perplexity CEO Aravind Srinivas praised Google DeepMind’s Gemini Neon with reasoning set to High for combining intelligence and speed; he said he is using it on Perplexity Computer and has migrated some of his teammates to it, a positive model-quality and adoption signal from a competitor’s CEO.
A user reported asking AI to clone them and create an explainer video for a new circuit board; Jason Calacanis praised the example as “Genius.”
Is Hinton Exaggerating, or Is LeCun Downplaying the Danger of AI?
- LeCun argues that text prediction alone does not give AI a model of the physical world. After leaving Meta at the end of 2025, his Paris-based AMI Labs pursued AI that learns physical-world models rather than predicting text and raised just over $1 billion in March.
- LeCun identifies concentration of power, including AI-mediated information diets, as a major concern; he argues that extinction warnings can help the largest labs win rules that restrict open research and entrench their lead.