We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
The math release gets organized pushback
The day after OpenAI published results from an unreleased frontier model, the response turned from awe to opposition. Exponential View put the release at 722 manuscripts in 372 families, averaging the equivalent of three hours of ChatGPT Pro thinking each. It noted that many results have been verified in Lean, but not all . The same post floated a split into "machine mathematics," verified but mostly read by AIs, and a smaller human "effective theory" of what people can understand .
Terence Tao reposted a statement from the Association for Human Mathematics . It urges mathematicians to "discontinue their work with OpenAI and to return to a vision of science that centers human understanding" . One commenter pointed out that the repost does not necessarily reflect Tao's own view .
The investor angle comes from Leo Polovets. He asks whether organizations will start "tokenmaxxing" research, for example a seed fund spending $50K or UPS spending $500M on tokens to find better routing algorithms . His reasoning: AI's R&D impact so far has been mostly on development rather than research. If compute can produce genuinely novel results, the right AI budget for some organizations could be much larger than they now assume .
A related data point: P0 Research reports that frontier models' "experimental research taste" has doubled roughly every three months since December 2025. It says Opus 5.5 now beats its expert human baseline, though most of those experts have not worked at a frontier lab .
Models: GPT-6 for everyone, cheaper Haiku
OpenAI is rolling out GPT-6 and "Intelligent UI" to all ChatGPT users. The new UI turns answers into interactive visuals and tools .
Anthropic released Claude Haiku 5.5, saying it costs about 75% less to run than Haiku 4.5 . Early reviews are mixed:
- Jerry Liu: for document parsing it costs about $1.20 per 1,000 pages and handles tables and reading order well for that price. It is weaker on charts, semantic formatting and bounding boxes .
- Bindu Reddy: calls it "MUCH worse than DeepSeek" and worse than Luna, but gives no benchmarks .
Separately, Anthropic's startup program offers up to $7,000 in Claude credits for a year . A Reddit commenter flags terms worth reading before applying. These reportedly allow Anthropic to develop competing technology, bar using the services to build competing products, and let Anthropic change the terms without notice. This is one commenter's reading of the terms, not verified here .
Power is the bottleneck: Texas freezes data-center permits
a16z reports that Texas has frozen new data-center permits. The large-load queue went from 63 GW to 474 GW in 18 months, more than 5x record peak demand. Only 9.5 GW is approved and about 4.3 GW is running . ERCOT has also paused approvals for data centers of 75 MW or more to switch on, including 17 that had finished every other step. The timeline is "TBD," and full grid connection can take 5–10 years .
Behind-the-meter power is a bridge rather than a replacement. SemiAnalysis tracks 75 GW of equipment orders for on-site power, but sites still plan to connect to the grid because on-site power almost always costs more . a16z also argues that data centers willing to cut usage at peak hours could add 100 GW to US grids without new plants . For scale, US grid additions planned this year total 86 GW: 43.4 GW of that is solar and none is new nuclear .
a16z's Ryan McEntush sees two ways for startups to break in: winning on integration or business model (Base Power), or bringing better technology (Heron Power's solid-state transformers) . His diligence point is that big firm orders rarely go to equipment without thousands of operating hours. So getting designed into a real project "matters far more than early deposits or revenue" .
The same day, a16z said Base Power raised $2.5B. Three years in, the company says it built a battery factory in 8 months, installs 100 batteries a day and powers more than 30,000 homes .
Agents move to the desktop, and gatekeepers resist
Sriram Krishnan described Microsoft's new Windows agent features :
- MXC containers: a local sandbox where agents run.
- Hybrid routing: actions can run on local models or be sent to cloud models.
- Copilot agents: one demo did taxes with no remote models.
- Hardware: a new generation starting with Surface Laptop Ultra and Nvidia RTX Spark.
Replit launched a Windows desktop preview that builds apps locally, with each build sandboxed using Microsoft Execution Containers and Nvidia OpenShell . Amjad Masad cites supply-chain attacks and catastrophic agent mistakes as the risks this design addresses .
On consumer agents, a guest on an Alexandr Wang–hosted interview described Muse's plan to bring agents to non-coders, whom the industry largely overlooked . Muse is free with subscription tiers and is exploring small fees on purchases or savings it facilitates rather than ads .
Paul Graham argues that Amazon banning agents is the first opening he has seen for a startup to build an Amazon competitor . More generally, he says any business that bans agents shows there is demand for a competitor that allows them .
Deals and funds
- HealthLeap: raised $38M across Seed and Series A from Sequoia, First Round and Hummingbird. It reports 39% more appropriate diagnoses with the same staff at Cedars-Sinai, and growth from 3 to 50+ hospitals in a year .
- Vitalize Care (YC W23): raised a $31M Series A for hospital staffing software. It says it manages millions of shifts a week across 15+ health systems, and that St. Luke's cut overtime and agency spend 54% in 12 weeks .
- Preference Model: a16z invested. The company builds RL environments for AI research and ML engineering for leading labs, and is open-sourcing Karotte, a framework it says was hardened over more than 1M evaluation runs .
- Sriram Krishnan: raising a $500M VC fund, per Axios via Dan Primack .
- a16z Japan: a16z opened a Japan office led by Hitoshi Yoshida, former president of Microsoft Japan and HP Japan. It is targeting AI, cybersecurity, robotics and defense .
Early-stage market read: Hustle Fund's September numbers
- Volume: Hustle Fund passed on 1,901 of 1,922 pitches, with 21 still in progress .
- Top pass reason: incorporation, at 47%. The firm wants Delaware C-corps or Canadian or Singaporean equivalents .
- Valuations: about 70% of founders sought valuations under $10M .
- Categories: 38% self-identified as AI/ML .
Tooling and science
- Perplexity: open-sourced pplx-embed-v2-late, multimodal embeddings in 9B and 0.6B sizes that share one embedding space. It enables OCR-free PDF search and querying on the device .
- OpenDocRouter: LlamaIndex launched a unified document-parsing API that serves models at cost plus a small transaction cut .
- OCR copying allegation: Vik Paruchuri alleged that Interfaze's new open-weight model copied and relicensed Chandra OCR 2 . This is an allegation, not established.
- AlphaGenome: DeepMind says it matched or beat the strongest comparison model on 25 of 26 variant-effect benchmarks . Its Atlas precomputes effects for about 9 billion single-nucleotide variants and is 30x the size of the AlphaFold database . Linking these molecular predictions to disease risk still needs more research .
- Angel Studios’ AI workflow includes vibe-coding product prototypes, using a collaborator to take promising experiments to MVP before handing them to the product team, and applying brand-guide harnesses across models; engineering adds harnesses when experiments break things. The team also built a tool that randomly serves different movie trailers to measure responses, and Harmon said AI speeds attrition and lifetime-value modeling when the data is clean.
- Harmon framed agentic shopping as a possible distribution opportunity: he said he was searching with Grock rather than Google and argued that businesses with APIs AI agents can use may reach shoppers who prefer buying through bots.
- Angel used a 19-minute skippable YouTube ad to pitch The Chosen; 17,000 people invested $11 million, with $800,000 in ad spend—described as just under 8% of the raise—and livestream pop-ups showing contributions as social proof. The roughly 3-million-member Angel Guild votes on films, can veto titles and gives feedback on rough cuts; Harmon said it had paid $300 million to filmmakers. As a financial caveat, the host cited a roughly $50 million loss the prior year; Harmon said marketing was a big expense and acquiring members was not cheap.
- The interview presents Muse as a breakout consumer launch: its opening touts 5 million downloads in 22 days, the host calls it the fastest-growing consumer tool since ChatGPT, and the guest says the team did not expect it to become a hit.
- Muse’s product strategy is to bring agent task completion beyond coders through a polished consumer experience; the guest reports use cases including cutting subscription costs, finding unused gift cards, navigating DMV and healthcare/medical-claims processes, and helping with small-business or social-media work. The longer-term ambition is a general-purpose assistant that helps people accomplish goals large and small.
- Muse is free and has subscription tiers; the guest says the team is exploring small fees on transactions where Muse helps users buy, save, or earn money, and is not spending much time on advertising currently.
- The guest says agent memory remains an unsolved problem and describes Muse’s current safeguards as a sandboxed secure VM, a sentinel monitoring information sent outside it, and user approval prompts for new websites or information sharing; a more confidential VM was planned, with possible speed trade-offs. He agrees that consumer trust can matter more than further increases in raw capability.
- On the broader AI ecosystem, he identifies safer, more generalizable algorithms as a promising direction with multiple startups working on it. He said broader AI-leader discussions were not happening in a private room at the time, while describing a White House accord for company-defined controls with internal, external-auditor, and board-level verification.
- Beyond consumer chatbots, Ng sees substantial enterprise opportunity in selecting and implementing AI workflows: banks may automate parts of loan underwriting and KYC, but identifying valuable, feasible uses and building reliable, tested, compliant, secure, private systems takes considerable work. He expects this implementation work to continue for at least a decade.
- AI-assisted coding is lowering the barrier to building custom software; Ng says software-engineering job postings are up and that AI-native engineers are scarce, while the required skill set has changed.
- Ng says open-weight models are widely used in developing countries to provide access without large budgets or dependence on continued outside access; they also let universities and local teams study models and adapt them to regional languages and use cases.
- Capability caveat: Ng says AI progress is faster on verifiable tasks such as coding, math and factual answers than on judgment-heavy work, which is also harder and more expensive to benchmark. He cites a task-based estimate that perhaps 30–40% of work in many jobs could be automated, while the remaining human work may become more valuable, increasing the need for upskilling.
- In education, Ng says chatbot use for homework can raise homework scores while lowering final-exam scores or measures of long-term learning; he is optimistic about AI learning products built on sound pedagogy and more personalized, one-to-one instruction instead.
- Ng argues that data centers are an important enabler of AI adoption and calls AI key economic infrastructure, making this a time to invest in technology.
- Texas’s large-load queue grew from 63 GW at the end of 2024 to 474 GW by June 2026, about 90% of it data centers and more than five times record peak demand. The state moved from directing data centers to pay for grid upgrades to freezing new permits; ERCOT also paused activation of data centers at least 75 MW pending an audit, including 17 projects that had completed its other steps.
- a16z presents peak-hour curtailment as a way for data centers to connect sooner, claiming U.S. grids could add 100 GW of such load without building a new plant. Separately, the accompanying article cites a Duke estimate that ERCOT could add about 10 GW without new generation if loads gave up 0.5% of annual grid electricity, setting transmission limits aside. ERCOT’s Batch Zero process offers large loads options to bring their own power or accept automatic curtailment when lines are full.
- Behind-the-meter power is a growing infrastructure market: SemiAnalysis is tracking 75 GW of equipment orders, although the article says sites generally intend to connect to the grid when available because on-site power usually costs more. The article identifies startup openings in integration and business models, exemplified by Base Power’s aggregated home batteries, and differentiated hardware, exemplified by Heron Power’s solid-state transformers. It cautions that first-phase adoption is difficult for new vendors: buyers favor equipment with thousands of operating hours, making design-in to a real project more important than early deposits or revenue.
- Texas froze new data-center permits after its large-load queue grew from 63 GW to 474 GW in 18 months—more than five times record peak demand. Only 9.5 GW was approved and about 4.3 GW was operating; a16z says the remaining requests include duplicates and speculative projects, some from developers who have never plugged in a GPU.
- Power availability is a deployment bottleneck: full grid connections can take 5–10 years, and ERCOT paused approvals for data centers of 75 MW or more to switch on—including 17 that had completed other steps—pending an audit, leaving the timeline TBD. Developers are pursuing behind-the-meter power; SemiAnalysis was tracking 75 GW of equipment orders, though a16z says the sites it knows generally still plan to connect to the grid because on-site power usually costs more.
- The potential buildout is large: a16z’s illustrative scenario of 100 GW running year-round equals 876 TWh, about one-fifth of 2025 U.S. electricity use; planned 2026 solar, wind, and gas additions would produce about 150 TWh annually at the cited capacity factors. The article argues that flexibility can help, but cannot eliminate the need for more generation and wires.
- Startup openings include integration/business models and differentiated grid hardware: Base Power’s fleet passed ERCOT pilot tests on its first attempt and was expanding to 50 MW; Heron Power’s solid-state transformers were set for a West Texas pilot with RWE. The source cautions that new power-equipment vendors face a hard qualification hurdle: large firm orders rarely come before thousands of operating hours, making design-in to a real project more consequential than early deposits or revenue.
- Texas illustrates a near-term AI infrastructure bottleneck: ERCOT’s large-load queue grew from 63 GW at the end of 2024 to 474 GW by June 2026, about 90% of it data centers; the state’s response escalated from requiring projects to pay for grid upgrades to freezing new data-center permits. ERCOT also paused energization approvals for data centers of 75 MW or more pending an audit, leaving the Batch Zero timeline “TBD” and the load forecast on hold; full grid connection can take 5–10 years.
- Developers are pursuing behind-the-meter power to accelerate time-to-power: the article says SemiAnalysis was tracking 75 GW of equipment orders for these assets, while noting that known projects generally still intend to connect to the grid because on-site power usually costs more. Flexible demand—such as cutting load or shifting computing—could help accommodate new data-center capacity.
- The analysis identifies power equipment as a startup opportunity through either integration and business-model innovation (such as aggregated home batteries) or differentiated technology (such as solid-state transformers). It cautions that shortage-driven “worse but faster” offerings may lose out if supply catches up, and that new vendors face a track-record hurdle: getting designed into a real project matters more than early deposits or revenue.
- a16z’s post cites a U.S. study finding that each 10% increase in data-center capacity corresponded to a 40-basis-point decrease in residential rates, reasoning that large customers can spread fixed grid costs across more electricity sales. The benefit depends on who pays for grid upgrades, however, and 56% of Texas voters say more data centers would raise their power bills.
- AlphaGenome predicts the effects of genomic variants. It operates at base-pair resolution across a context window of up to one million base pairs; the interview reports it matched or beat the strongest comparison model on 25 of 26 variant-effect benchmarks.
- AlphaGenome Atlas precomputes molecular-effect predictions for roughly 9 billion possible single-nucleotide variants in a petabyte-scale dataset, described as 30 times the size of the AlphaFold database. The Atlas is made available to the scientific community; AlphaGenome model weights are available for academic research and fine-tuning.
- A key translational caveat is that AlphaGenome predicts molecular phenotypes; connecting these to whole-organism traits or disease susceptibility still requires further research. The team also identifies cell-context awareness, more training data, and linking molecular predictions to organism-level effects as ongoing work.
Paul Graham says Replit is more important than he had previously thought: AI-powered apps can make organizations somewhat aligned with AI, but he sees code as the route to the deepest organizational change.
- Paul Graham argues that Amazon’s ban on agents creates an opening for an Amazon competitor: he expects agents to become a major way people buy goods and says shoppers may not want an Amazon-supplied agent doing it for them.
- More broadly, he treats companies’ agent bans as a startup opportunity, reasoning that restrictions imply users want agents and may create demand for competitors that allow them; a nested reply sums up the thesis as “Your usage restrictions are my opportunity.”
- Texas is a major AI-infrastructure bottleneck: ERCOT’s large-load queue grew from 63 GW at the end of 2024 to 474 GW by June—more than five times record peak demand, with about 90% attributed to data centers. Speculative multi-site applications and concern that ratepayers could fund upgrades for projects that never materialize contributed to a pause that escalated to a freeze on new data-center permits. ERCOT also paused approvals for data centers of 75 MW or more to switch on pending an audit; 17 projects that had completed other ERCOT steps were caught in the pause, with timing listed as “TBD.” Full grid connection can take 5–10 years.
- Developers are turning to behind-the-meter power to reduce time-to-power, but the article says sites generally still plan to connect to the grid because on-site power almost always costs more; islanded power also brings reliability and fuel-supply challenges.
- Startup opportunities include grid flexibility and differentiated power hardware: a Base Power fleet with GVEC passed ERCOT pilot tests to sell into the wholesale market and is expanding to 50 MW, while Heron Power is set to install solid-state transformers at a West Texas battery site with RWE. The article identifies integration/business-model innovation and new technology as startup entry points, but cautions that commercialization depends on getting products designed into real projects and proving reliability, competitive pricing, and service; large firm orders for unproven equipment are difficult to secure without substantial operating history.
- a16z says Base Power has raised $2.5B to build a nationwide power company; three years in, it says the startup built a home-battery factory in eight months, installs 100 batteries a day, and powers more than 30,000 homes.
- Base was founded in 2023 by Zach Dell, who worked at Blackstone and Thrive Capital, and Justin Lopas, whose experience includes manufacturing at Anduril and building rockets at SpaceX. Its residential-battery approach uses homes’ existing grid connections to bypass interconnection queues and transmission congestion.
- Base Core is a 39.2-kWh home battery; in markets where customers can choose providers, Base offers electricity for three years at a fixed rate plus delivery fees, with typical bill savings of 10–20%, and uses the batteries to trade power with the grid. It charges batteries when power is cheap and sells it back when prices are higher.
- Utility partnerships rose from less than 5% to more than half of Base’s sales volume in a year; Austin Energy contracted for 40 MW of home batteries and CoServ for 100 MW.
- U.S. grid additions planned for this year total 86 GW: 43.4 GW solar, 24.3 GW batteries, 11.8 GW wind, 6.3 GW gas and no new nuclear; AI labs are planning for hundreds of gigawatts. Solar is the only source already produced at that scale, but most production is in China, while scaling firm non-gas power is more complicated.
- Texas shows the deployment bottleneck: ERCOT’s large-customer queue grew from 63 GW at the end of 2024 to 474 GW by June, about 90% of it data centers; the source says many applications are speculative and lack customers. Texas has frozen new data-center permits, and ERCOT paused approvals for projects of 75 MW or more pending an audit, including 17 that had completed other ERCOT steps; connecting a large data center can take 5–10 years. Behind-the-meter equipment orders offer a bridge, but sites generally still aim to connect to the grid because on-site power is usually more expensive.
- The shortage creates startup openings in integrating existing hardware and business models, such as Base Power’s aggregated home batteries, or in new technology, such as Heron Power’s solid-state transformers. For investors, the source flags a cold-start risk: project design-in matters more than early deposits or revenue, and new vendors must prove reliability, competitive pricing and service capability.
OpenAI is rolling out GPT-6 and Intelligent UI in ChatGPT for everyone; the UI provides fast, interactive answers, makes everyday questions more visual and complex topics easier to grasp, and offers task-specific interactive tools on the spot. Sam Altman called the change long-awaited and said he would hate to return to the old version of ChatGPT.
Paul Graham says successful founders he has spoken with wish they had hired fewer senior managers from outside, while he has not encountered one who wished they had promoted fewer people internally . One founder’s explanation was that external hires may be a different kind of person, with loyalty to their career rather than the company .
Scott Kupor joked that OPM was coming after Sam Corcos to take his title as the government’s biggest USTechForce employer . In the linked post, Corcos said they still need more engineers and invited interested people to DM him .
Dan Primack reported that @sriramk is raising a $500 million new VC fund . Scott Kupor praised him and asked for a running total of AUM for the “a16z mafia” .
Scott Kupor highlighted a Washington Post perspective that “SI” is generating so many new jobs that a slowdown is a concern; his post does not clarify what “SI” means, so it is not an unambiguous AI-labor signal.
Scott Kupor highlights a historical shift in tech talent: women made up as much as 40% of the early programming workforce, while female employment in programming peaked in the 1980s and has not returned to that level.
Tech Force held an event with OPM Director Scott Kupor, tech founder Joe Lonsdale, and Under Secretary of War Michael on how technologists can serve America, advance “SI,” and build tools to strengthen national security; the post gives no further details on specific programs or tools.
Vitalize Care (YC W23), which is building an AI operating system for hospital staffing, scheduling, and capacity, raised a $31M Series A. The company says it manages millions of shifts per week across 15+ health systems and saves clinical leaders over 2,000 hours daily; St. Luke’s Health reduced overtime and agency labor spend by 54% within 12 weeks.
