We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Meta is back at the frontier, depending on whose evals you trust
Alexandr Wang launched Muse Wonder as Meta's "largest and most capable model yet." He called it state of the art for personal and professional agents . In an interview, Meta said it focused on memory and on the model's ability to keep working through complicated tasks . It said the lab was rebuilt after it was "not happy with the trajectory" of Llama 4 . It also expects the time between successively larger model classes to "compress" . The compute behind this is large: the Ohio cluster is scaling past 1 GW, and the Louisiana cluster is planned to reach 5 GW over the next couple of years .
Early outside views disagree:
- Positive: Aravind Srinivas says Meta's model compares favorably to Fable 5.1 on WANDR evals . Gavin Baker says Muse and Meta are now "significantly ahead" of Gemini on capability .
- Skeptical: Bindu Reddy first bet the model was "benchmaxxed" . She later said it is "nowhere near Fable or Astra level," and Abacus plans to publish its own benchmark .
- Usage: Muse's D7 retention is above 30%, higher than ChatGPT's. Its D14 retention of 22% is just below ChatGPT's and well under social apps such as WhatsApp (67%) .
Why it matters: a fourth well-funded frontier lab competing for personal agents puts pressure on agent-app startups that rely on any one model. Until independent benchmarks come in, capability claims are contested.
Recursive self-improvement gets yardsticks and a stricter definition
Three items point to RSI moving from slogan to something people can measure:
- TasteVal (P-Zero Research) tests experimental research taste against a baseline from 24 experts who recently worked at labs such as OpenAI and DeepMind . It reports that frontier models' taste has doubled every 3.0 months since December 2025. It says Opus 5.5 beats the human baseline at about 1/30 of the experts' cost per attempt . Import AI's author now thinks full RSI is more likely in 2027 than 2028 .
- DeepRSI has agents work for 48 hours on ML tasks taken from ongoing proprietary research. Existing benchmarks top out around 12 hours . Garry Tan calls the next 18 months RSI's "Kitty Hawk moment" and says "evals will be how you can tell" .
- Anthropic's internal index says Claude led 26% of measured model R&D in August, with humans supervising. None of the measured work was fully autonomous .
On diligence, a16z's Martin Casado and MakerMaker.AI founders Dhaval Adjodah and Owen He say many "RSI" pitches are just people using AI to build better AI . Their bar is that the model improves its own weights, harness and task prompt, and that the gains hold up once experiment costs are counted. They track spend on a "bleeder board" and assign one engineer to try to break their results. They report early signs of the system getting better and cheaper at the same time . Investors can use the same questions on RSI pitches: what exactly improves, and at what total cost.
Funding: Series A/B is hard, so get profitable or keep your pro rata
Rohit Mittal describes the trap. Vertical AI SaaS companies reach $1–3M ARR in two years, then find "almost no interest" from investors at 1–3x growth. They are left with large preference stacks, burn they can't sustain, and exits that leave little for the team . Garry Tan's answer: run profitably with "maximal agentic loops," because $10M of profit reinvested each year works like "an annual Series A no dilution" . Martin Casado warns VCs against getting hung up on TAM, revenue quality and margins early in new technology waves .
Hunter Walk expects more pro rata conflicts. Early-stage funds are now big enough to want follow-ons across several rounds. Rounds come faster. Hot secondaries make pro rata feel tradable. Fund-of-funds and SPVs now exist at scale to finance it . A widely shared example: Initialized Fund 1 returned 55x. The first 98 companies produced about 1x DPI, and Instacart, where the fund could not take its pro rata, added another 1x. Coinbase, where Tan kept taking pro rata, produced the final 53x . Separately, Atlassian's Mike Cannon-Brookes says raising $15M at pre-seed has become too easy, and blames VCs for it .
Agents as customers
Sierra's Personal Agent Protocol published its first draft and announced 35 new design partners. The founding group includes Meta, Shopify, Stripe and Walmart. The protocol aims to let companies know when they are dealing with an agent and whom it represents . Bank of America is a design partner . Sarah Guo reports that a Sierra Summit executive said "friction had been a very good business model" and their company would have to change. Her point is that agents threaten revenue that depends on customer inattention . On the founder side, Paul Graham met a founder who declined a sales cofounder because "most of his sales would be to agents," and Graham admitted he could be right .
Anthropic launched dynamic workflows for Claude Managed Agents. Claude now plans the work itself and writes the program that coordinates its subagents, so developers no longer configure the roster . The model is not told its budget while it runs, but developers can set a session cap .
Perplexity's numbers
According to an interview recap, Perplexity is valued at about $20B, with about $2B raised and $750M in annualized revenue as of August, growing 5x year over year . Srinivas cited a one-person company that makes $3M a year and spends more than $1M of it on Perplexity Computer . Perplexity says its agents commit to a shared Git repo with under 2% collisions. It also says plain ls/grep beat semantic search on answer accuracy, 63% vs 37%, in financial-document analysis . That is a useful data point against assuming vector search is required.
Infrastructure and models
- Microsoft Maia: starting next year, Microsoft will scale its Maia chips to multi-gigawatt deployments as the primary inference system for its own products. It claims Maia 300 delivers 40% more tokens per watt than leading GPU systems .
- Distillation: Anthropic says it has cut the data harvested from Claude through distillation attacks by about 90–99%. David Sacks presents this as evidence that labs, not regulators, should handle distillation. He notes that the administration did not impose the open-model restrictions that were proposed .
- Western open weights still trail Chinese models: Reflection's 501B-parameter Beam scores 44.4 on DeepSWE vs 61 for GLM-5.3. Reflection says Beam matches GLM-5.2 on reasoning with 3–4x less compute . Mistral Large 4, tuned for cyber tasks, may be the strongest Western open model at release but still trails Kimi K3 and GLM-5.3 .
- Enterprise model provenance: Atlassian runs Nemotron, Mistral and other open models but no Chinese models in production. Cannon-Brookes says most large companies that claim to have "built their own model" have fine-tuned someone else's .
Deals and programs
- HyLight (YC S23) raised $14.5M for autonomous, hydrogen-powered airships that inspect power lines, pipelines and railways. Four are flying missions in France, with 15 planned to operate at once by 2027 .
- a16z Speedrun SR008 is taking applications through November 1. It invests up to $1M plus more than $10M in partner credits, and runs February 1 to April 16, 2027, in San Francisco .
- Avalanche Energy signed a letter of intent with the Tri-City Development Council for its first compact-fusion manufacturing plant. The plant is estimated to cost close to $1B in private funds .
- Andrew Ng is bullish on databases and core infrastructure: he expects agentic workloads to drive more demand for compute and storage alongside AI training and inference, and says builders are already reporting compute shortages.
- DeepLearning.AI’s AI-engineering skills map draws on more than 10,000 job descriptions, dozens of structured interviews, and surveys; it highlights building and deploying AI applications, software fundamentals, coding-agent use, and shaping what and how to build. Ng says software-engineering job postings are up and his teams cannot find enough skilled AI engineers.
- AI assistance is broadening developers’ work from implementation toward product and architecture decisions. Ng recommends rapid, disposable prototypes and learning in early-stage projects, with more careful attention to architecture, customer requirements, and security as project needs and risks change.
- Ng warns that cognitive offloading to LMs can increase work completed while reducing learning, citing emerging studies on homework and exam performance. He described recently announced Learn Vector as a “frontier human development lab” pursuing more personalized, one-to-one learning; products were not yet announced, though he said internal prototypes were promising.
- The program’s opening says Nvidia was reportedly weighing an investment in Perplexity at a valuation above $30 billion; it is presented as a report, not a completed financing.
- Perplexity says its revenue is up 5× year over year as it evolved from an answer engine into search and browser agents, and says self-serve enterprise growth accelerated as the product began supporting business workflows. It cited a one-person newsletter-automation business earning $3 million a year and spending upwards of $1 million a year on Perplexity’s product.
- Perplexity’s agent strategy is to orchestrate web search, browser access, personal context, files and tools across models; the company argues that AI products can create value through deep, recursive usage rather than relying only on network effects. Its CEO also argued that future U.S. growth could come increasingly from smaller and mid-market AI-native businesses that are more productive and capital- and headcount-efficient.
- Perplexity says it has no plans to build its own data centers and instead aims to orchestrate between cloud and local hardware. It argues local compute becomes more attractive for long-context, personal-agent workloads, and describes a fully local portable-computer effort with Nvidia and a cloud-to-local option with Apple; it reports interest in hybrid deployment from finance, legal and healthcare.
- On agent security, Perplexity says it is working on open-source Bumblebee for devices and Numbat for servers to audit agent activity and flag sensitive-data access; its CEO warns that single-agent testing may miss risks from collaborating multi-agent systems.
- Perplexity’s longer-term research ambition is agents that build prediction models, run real-world experiments and test hypotheses; it acknowledges current frontier models lack persistent, parallel autonomy, especially for questions where truth is hard to verify.
- Atlassian CEO Mike Cannon-Brookes says agents will automate some functional tasks but can also move enterprise collaboration work forward; he expects horizontal and vertical agents to coexist with application-specific interfaces. He argues that abandoning the user experience is a risky bet, and identifies design and an enterprise context graph as strategic priorities.
- Atlassian describes its teamwork graph as an index of company-specific decisions, documents, data, people, and code. Its view is that organizational relationships must stay current while historical information is retained for its learnings, rather than treated as equally useful current facts.
- The case for specialized-model companies remains unproven in Cannon-Brookes’s view: proprietary data or genuinely specialized applications such as robotics can justify specialization, but models must sustain price/performance gains over improving foundation models. He says many “own model” claims are actually open-weight models fine-tuned for a use case, not models trained from scratch.
- Enterprise AI buyers increasingly care about cost and price/performance; Cannon-Brookes points to model routing, blending, and harnesses as ways to optimize them. He says large technology companies may build routing internally while smaller or less technical firms can use third-party gateways; he also expects inference costs to fall substantially even as cheaper inference increases usage, with technology production currently outpacing customer consumption.
- On model sourcing, he reports some customer pushback to Chinese models but not to open models; Atlassian uses open models and says it does not use Chinese models in production.
- On early-stage founder durability, Cannon-Brookes worries that it has become too easy to raise $15 million at pre-seed, hire, and then exit quickly; he links sustained commitment to enjoying product creation and solving hard problems, not just pursuing the latest technology wave.
- The startup’s interviewed team includes a speaker with a physics background and a PhD in reinforcement learning from MIT Media Lab, and Owen, whose PhD work was in continual and meta-learning and who worked as a research scientist at DeepMind. They distinguish recursive self-improvement (RSI)—AI improving parts of itself—from human-led AI-assisted development, and describe a hybrid multi-agent system that includes human researchers and engineers and aims to co-optimize interdependent layers from prompts and harnesses to model weights and chips.
- The team is targeting inference efficiency, arguing that improvements can both reduce resource costs and enable more inference; it says its agents choose what to optimize. The founders report early “glimmers” of getting better results with less energy, with results growing over recent months.
- The founders say many papers claiming RSI have failed their internal tests at scale, and stress empirical measurement because results can be affected by noise and reward hacking; they also say they have held back releasing results until confident. They argue that improvement driven by ever-increasing resources is not RSI and recommend accounting for experiment costs and human hours, not just compute.
- The guest argued that quantum computers are not broadly faster computers: he described current systems as low-clock-speed and single-core/threaded, with no quantum RAM, and said their standout speedup is limited to factoring and elliptic-curve discrete-log problems—leaving a narrow business case. He also said he had lost millions shorting quantum-computing companies, a caution on timing the thesis.
- He sees photonic computing as a potential alternative to GPU scaling: optical matrix-multiplication cores could outperform Nvidia, and he said his group was investigating a possible company, while flagging unresolved nonlinearity and storage issues. He expects current GPU generations could eventually be displaced, but said he is long semiconductors now; efficiency gains could pressure Nvidia, while Jevons-driven demand might offset them.
- He sees AI demand extending beyond knowledge-worker labor, with AI-founded businesses across industries; he cited Core Automation, an under-the-radar startup from a former OpenAI colleague whom he said invented OpenAI’s reasoning model. He said he would buy OpenAI and Anthropic, has OpenAI exposure, and views Meta’s distribution and product work on Muse favorably.
- As a counterpoint to AI-biotech optimism, he predicted there would be no AI-relevant drug in the next 10 years and argued AI currently contributes little to drug discovery, citing his experience building a Cray cluster for drug analysis and later using GPUs.
- US adult use of GLP-1s for weight loss rose from 3% to 11%; Exponential View estimates end-2025 GLP-1 sales at $76 billion versus $63 billion for generative AI, calling it a possible revenue crossover.
- Pricing is shifting rapidly: Novo’s US semaglutide self-pay price fell from $499/month to as low as $199/month for the first two doses, with an average price of $350/month; India’s roughly $13/month generic is a lower dose and not directly comparable. Lilly reported a 13% realized price cut alongside a 60% volume boom, attributing most of the price decline to China’s reimbursement list and lower US cash prices. Access remains uneven: four US Medicaid programs dropped obesity coverage, and semaglutide and tirzepatide patents are not due to expire until 2032 and 2036, respectively.
- Weekly injectable weight-loss efficacy has improved by 10 percentage points from semaglutide in 2021 to retatrutide results reported in May; oral GLP-1s are more convenient and easier to manufacture but have been less effective than injections. Persistence is a commercial caveat: one Danish study found half of users stopped within a year, while UK persistence was 32% at one year and 15% at two years.
- Potential spillovers beyond weight loss include a study in adults with obesity and alcohol use disorder in which semaglutide plus talking therapy cut heavy-drinking days by 41%, versus 26% with therapy alone. Evidence of broader economic savings is mixed: a Danish study found 17% less long-term sick leave over four years, while another found no insurer medical-cost savings over five years and $0.22 more in medical spending per dollar spent on GLP-1s. UK users are also shifting grocery spending away from chocolate toward gum and mouthwash, a consumer-market signal the source connects with Hershey’s push into gum and mints.
- In a sponsored Box segment, CISO Heather Ceylan frames AI risk as capability × access × autonomy × scale, and argues that guardrails should be supplemented by audit trails, human approval for consequential actions, and limited permissions.
- Meta positioned Muse Wonder as a larger frontier model for personal and professional agents, emphasizing active memory across days or months, persistence on complex tasks, and improved coding; Meta said it would initially offer Wonder mainly through paid Muse subscriptions because it costs more to serve, with API, OpenRouter, and Muse Code access also planned.
- Meta said it rebuilt its AI lab after dissatisfaction with Llama 4’s trajectory and pursued a small, talent-dense team with concentrated compute; the research group was described as fewer than 200 people, closer to 100. The speakers estimated that operating a gigawatt-scale cluster costs about $15 billion annually, with roughly $50–60 billion in upfront capital expenditure, and said tens of billions of dollars is a reasonable estimate for frontier-model development; Meta cited its Ohio cluster scaling above 1 GW and a Louisiana cluster planned to scale to 5 GW.
- Meta executives described agent-led shopping as an emerging commerce channel: agents can use browsers as a fallback, while direct business integrations should make transactions faster and more efficient; they expect agentic commerce to grow and identified subscriptions, commerce-related revenue, and enterprise services as potential monetization routes.
Shaan Puri recounts that at 21 he and two cofounders pursued a sushi-chain idea despite having little relevant food experience, spending months on plans, forecasts, and operating details before putting the business into action. A Noodles & Company founder advised them to secure a location, open within 90 days, serve customers, and learn how to make them happy; Puri says they shifted from planning to action, and a later mentor praised their unusually high bias to action.
At a creator Q&A at OpenAI Dev Day, Sam Altman identified iterating quickly on ideas as an increasingly important skill for using AI. He said there may be more jobs, but they will be different, and predicted that AI could operate computers 20 times faster than people “very soon”; he noted that current interfaces were built for people and have inherent slowness.
Garry Tan said he would bet Muse will win, linking a post that identifies it as Meta Muse and describes it as one of the highest-retaining AI products. The post reports D7 retention above 30% (higher than ChatGPT) and D14 retention of 22% (slightly below ChatGPT), while noting that D14 is below the cited social-app benchmarks: WA 67%, YT 59%, Instagram 54%, and TikTok/Snap 37%.
Garry Tan endorsed pro rata as powerful for investors . The linked account says his first fund returned 55x, with about 1x DPI from its first 98 companies and another 1x from Instacart, where it could not take pro rata; it attributes the final 53x to Coinbase, where Tan kept taking pro rata through multiple rounds .
A purported reporter’s outreach was identified as a phishing attempt after the recipient found the supposed journalist unfamiliar; the post links an OSINT write-up for further details.
- Garry Tan recommends that startups prioritize profitability and run with “maximal agentic loops,” arguing that reinvesting $10M in annual profit can function like a nondilutive Series A and that leading startups are already doing this .
- The financing pressure behind this advice: Rohit Mittal says vertical AI SaaS companies reaching $1M–$3M ARR in two years or $5M–$10M in three to four years can still find little Series A/B investor interest at 1x–3x growth; existing preference stacks and burn can leave them neither sustainable nor venture-fundable, while exits may offer little to teams .
Garry Tan says the next 18 months could be a “Kitty Hawk moment” for recursive self-improvement (RSI), and that evaluations will show whether it is happening. His quoted post presents DeepRSI as a new standard for measuring RSI: unlike existing short-horizon, contributor-based benchmarks, it has agents work for 48 hours on ML tasks derived from ongoing proprietary research.
Paul Graham said AI’s trajectory surprised him: instead of starting with a “perfect fly” and developing toward human capability, it first seemed human-like but unreliable, then began moving toward perfection.
Paul Graham recounts a founder rejecting the suggestion to add a more outgoing sales cofounder because he expected most sales to be to agents; Graham conceded the founder could be right. This is an early-stage founder’s hypothesis that agent customers could change sales-team needs, not evidence that the model is established.
- Sierra, Instinct and Meta announced the proposed Personal Agent Protocol, a standard for companies to recognize and work with customers’ agents; Bank of America joined as a design partner.
- Guo’s thesis is that agents can turn effort-abandoned tasks into transactions and erode revenue built on customer inattention: Instinct has already handled some users’ medical-billing errors end-to-end; she imagines agents shopping rates and removing setup friction that delays new-phone purchases, while Instinct’s founder expects easier ordering to bring delivery services more business. As a precedent for latent demand, a randomized trial of information and application help—not agents—tripled food-stamp enrollment among likely-eligible older adults from 6% to 18% in nine months; Guo expects agentic software to increase government outlays.
- For Baseten’s underwriting of roughly 150 compute suppliers, email follow-up and extracting numbers are easier to automate than verifying supplier claims and delivery capacity, which can still require on-site checks or validation through suppliers’ suppliers.
Garry Tan argues that foundation-model companies have two winning positions: being the absolute best or being the best at the lowest cost, a quality-versus-cost lens that excludes undifferentiated middle positioning from his thesis.
MakerMaker.AI co-founders Dhaval Adjodah and Owen He discussed recursive self-improvement (RSI) with a16z’s Martin Casado; He says many pitches labeled RSI are instead humans using AI to build better AI, while the key distinction is whether the model itself is being optimized. Their bar for RSI is that the model improve its own weights, harness, or prompt, and that gains hold up after experiment costs are counted; the team says it tracks experiment spend and has an engineer try to break results. They report early signs, over the past few months, of their system getting better and cheaper.
Speedrun opened applications for SR008 through November 1, 2026. The post’s introduction calls it a January 2027 start, while its program details give dates of February 1–April 16, 2027, at a new Jackson Square office in San Francisco. The program offers up to $1 million in investment and more than $10 million in credits from 200+ partners, plus investor access at Demo Day. Founder support includes growth, distribution, and fundraising help, customer introductions through a network of 150+ C-suite founding members and 1,000+ executives and operators, and access to a 171,000+ person talent network.
