We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
1. Funding & Deals
Event Horizon Labs is an early-stage bet on self-improving domain expertise. Dalton Caldwell says he led the Series A in EHL, a new quantitative research lab founded by Owen. Its thesis is to automate the type of work Owen previously did at Citadel with self-improving AI; Caldwell calls Owen one of YC’s most talented AI researchers and says several teams are pursuing the same direction.
2. Emerging Teams
Suhail’s autonomous-AI-scientist project is moving from experiment to compute-backed team. The build log says the project has a validated basic RLVR post-training stack, made its first hire while recruiting for post-training or low-level optimization, and grew from one person to three. After a key research component worked but needed scaling—and GPUs were delayed by networking issues—Suhail reported that much greater quantities of compute were locked down and ready.
The diligence question is whether the validated post-training loop compounds once the new infrastructure is online; the founder’s own updates frame scaling research and compute access as the immediate bottleneck.
3. AI & Tech Breakthroughs
Agent training is becoming an environment-and-verifier business. A current industry talk describes the shift from low-skilled crowdsourced behavior-cloning data toward expert-built environments containing realistic “worlds,” high-fidelity app clones, and tasks with rubric or unit-test verifiers. The speaker reports 2.5 million expert hours in the second quarter; in one 1,800-task post-training run using about $500,000 of compute, the reported score rose from 4.7% to 26% and generalized to other benchmarks.
That points to a valuable infrastructure layer beyond raw model supply: humans still have to measure performance beyond a model’s frontier, while the next data wave is moving toward 100–1,000-hour tasks and social interaction—an area where the speaker says only about 1% of evaluations currently measure performance.
Robotics reliability is being attacked with reinforcement learning and memory, not demos alone. The robotics talk reports a general-purpose value function trained on robot experience, with human intervention used to avoid dead-end trajectories; the resulting policy ran for 13 hours, exceeded 90% espresso success, and gained roughly 2x box-building throughput from the RL stage. Multi-timescale memory—short video memory plus compressed text for longer history—enabled a non-repetitive kitchen-cleaning task lasting 10–15 minutes, while a single Pi07 model reportedly matched or outperformed fine-tuned specialists and generalized to scarcely represented appliances and a new robot platform.
4. Market Signals
The near-term moat is shifting from the base model to post-training, domain data, and cost control. The post-training thesis is to use frontier models to reach product-market fit, then use product data to encode a company’s distinctive taste and expertise. Because applications can be cloned from screenshots, the tuned model becomes the proposed defensible asset; the speaker says post-training can also reduce serving cost by 5–10x. The same talk warns that reward hacking and training-to-serving drift can make apparent progress useless, making repeatable evals and production A/B tests part of the product rather than a launch afterthought.
Vertical agents are being framed around cash outcomes rather than task volume. Stuut’s collections analysis covers billions of dollars of receivables and finds that the largest 10% of past-due invoices hold 65.7% of overdue dollars. In its data, 81.7% of outbound collection emails require no human involvement, while escalations, disputes, manual calls, and broken-promise follow-ups remain human-resolved; agent-run teams average 1.35 outbound asks per $1,000 collected, with three in five completed tasks resolving fully automatically. The product pattern is clear: automate the search and repetitive asking, while reserving people for relationships and exceptions.
AI is changing founder leverage faster than it is changing the fundamentals of company building. Garry Tan argues that agentic and vibe coding can make one person “400 of that person” and expects a wave of experienced 35–45-year-old technical founders. Paul Graham’s counterweight is that almost all standard startup advice still holds: the core remains building what users need and finding growth. The investment implication is to widen founder sourcing toward experienced operators without relaxing scrutiny on distribution and durable demand.
Local and open inference is becoming a meaningful deployment channel. Hugging Face says Transformers.js crossed 10 million monthly downloads—nearly 10x its level six months earlier—and attributes local adoption to free, private execution amid compute shortages and cyber-attack risk. Cohere’s North Micro Vision adds to the supply side: its smallest vision-language model is open-source under Apache 2.0 for document understanding.
5. Worth Your Time
- Watch RL Environments Explained: How AI Agents Learn Real-World Work. The clearest explanation in the corpus of why agent progress depends on realistic apps, expert-built tasks, and verifiers—not just more model-generated trajectories.
- Watch Chelsea Finn: This is the State of the Art in Robotics. Focus on the memory and generalist-policy sections for a concrete view of how robotics is moving from isolated skills toward longer, compositional tasks.
- Watch Post-Training Is How You Keep Your Taste. Useful for deciding when a startup should rent frontier intelligence and when its product data, evaluation loop, and economics justify owning a specialized model.
- Read Paul Graham on startup advice in the AI era. A short corrective to claims that AI has already overturned the basic work of finding a product users want.
- Gary (YC partner) says agentic/vibe coding changes founding-team math: a single founder can now effectively be '400 of that person', though co-founders remain net-positive all else equal .
- Gary says AI-agent workflows now let companies go from 0 to ~$15M ARR in about four months with 2-3 people and a few hundred agent 'skill files'; he advises founders to 'skillify' tasks into markdown + code + tests runnable as cron jobs — 'a markdown file is an employee' .
- On SaaS, Gary says it's 'not totally clear' pure per-seat SaaS exists in another 5-10 years; it's fine as a wedge but must leap to a moat around data or network effects, and the old 10-20x next-12-month revenue multiple is gone .
- He sees a mega-trend of 35-45-year-old founders with operating experience (e.g., Peter Steinberger) who, amplified by agents, can outperform an entire big-tech department; young Patrick Collison-type founders will still emerge .
- Gary has been building and using agent tooling (OpenClaw, gbrain, Garyslist.org) and says 'token maxing' — full-strength agent usage — costs ~$50-100K/year but lets founders/CEOs 'live in 2028' .
- Enterprise adoption example: Brex CEO Pedro built an open-source safety layer (Crabtrap) around OpenClaw to monitor all agent network traffic in a regulated fintech, and uses agent-reviewed meeting transcripts to surface conflicts two levels down .
- YC-funded Clawvisor released a bug-report endpoint built for agents, not humans; an LLM replies in real time that the bug is filed, will be fixed, and suggests a workaround .
- Consumer AI: Gary expects the next computer to be voice + memory, predicts 2027 will be the 'harness wars,' and says when today's frontier-level compute costs $50-100 in 2-3 years, there will be a war for a billion consumers; the interviewer adds that token prices for recently-frontier models are collapsing to zero, unlocking consumer AI with zero-cost distribution .
- Timeline caution: Gary argues AI disruption will take ~20 years because bureaucracy and human limits slow every institution; incumbents like Microsoft have structural moats, and the AI-native generation (7,000 mostly 18-22-year-olds at YC Startup School) will drive the eventual shift .
Physical Intelligence's founder (company founded ~2 years ago to let any robot do any task in the real world) argues physical AI must operate fully autonomously and make far fewer mistakes than prior ML systems because its decisions affect the physical world; he cites Waymo surpassing 250k weekly autonomous rides a year ago as evidence this is achievable .
Physical Intelligence built a scalable RL post-training recipe for high-reliability manipulation: human interventions to avoid dead-end trajectories plus a general-purpose value function trained on robot experience to estimate progress/time-to-success, which cuts the number of attempts needed. Results included a 13-hour continuous latte run and >90% espresso success, and ~2x throughput from the RL stage alone on Dandelion Chocolate Factory box-building .
Because most robot foundation models lack memory and naive context is expensive (10s of multi-camera video would be ~500k tokens, or 10k tokens even at 1fps), Physical Intelligence added multi-timescale memory: ~10s video memory plus compressed textual summaries for longer horizons, enabling a 10-15 minute fully autonomous kitchen-cleaning task .
The single general-purpose Pi07 model - trained on heterogeneous data (robot demos, policy rollouts, human videos, web data) with memory, instruction, subtask, metadata, and optional subgoal-image prompting - matches or outperforms per-task fine-tuned specialist Pi06 models (RL-post-trained and SFT) out of the box .
Pi07 shows compositional generalization: operating a barely-represented air fryer (~3 training episodes) and folding clothes on a Bioarm UR5E platform with zero folding data on that robot, with held-out folding performance approaching human teleop; ablations show diverse data and metadata prompting are key, with metadata letting the model benefit from added low-quality data .
Physical Intelligence says physical intelligence has reached the right side of the generalist-AI timeline (analogous to the GPT/DALL-E era) and is being deployed via YC companies Ultra (laundry folding) and Weave (warehouse packaging); the model family adapts to bimanual arms, drones/quadcopters, surgical robots, and tractors .
PI Zero and PI Zero 5 are open source. The founder advises small teams to start from a generalist policy and fine-tune it, except in constrained deployments (e.g., surgical robots in ORs with no internet and weak GPUs); he expects the robotics open-source path may not exactly mirror LLMs given embodied data/hardware costs, but is optimistic a strong open-source community will exist .
On market timing, he expects no single ChatGPT-like distribution moment for robotics because physical hardware deployment is slower, but ChatGPT-level robot capabilities are on the horizon within the next few years .
Speed is improving via RL (a release called
RL Tokenshowed policies faster than human teleoperation), and he highlighted an emergent left/right hand transfer during pinwheel assembly that was never present in training data .
David Friedberg argues China's AI advance now endangers the core of the US economy: with energy output heading toward ~8 TW versus ~1 TW in the US, 20x the automated manufacturing square footage, and open-sourcing of AI, China is positioned to commoditize not just manufacturing but knowledge services and IP, which have been the heart of US economic strength for 30 years — a competitive dynamic relevant to AI-moat theses .
Rahm Emanuel argues the US innovation system — new technology entrepreneurs, venture capital, and the DARPA research arm — can deliver within ~5 years, as long as legacy defense incumbents (Lockheed, Raytheon, General Dynamics) and Pentagon procurement don't undermine it .
Emanuel flags an AI-infrastructure labor bottleneck: the US is ~500–600k electricians short for data centers and 150k short for broadband .
On research funding, Emanuel calls NIH, NSF, DARPA, national labs, and universities the US' "crown jewel," criticizes repeated proposed NIH budget cuts (40% in each Trump budget, resisted by Republicans down to ~15%), and proposes a 10% levy on prediction markets and online sports betting to double US research funding through NIH/NSF/DARPA .
- Mercur, an RL-environment/agentic-data vendor, appears to have scaled to a ~$2B revenue run rate in four months, per a host remark . It became the primary agentic-data vendor to frontier labs and app-layer companies (Harvey, Cira, Cognition, Ramp) after its first big project, deep research .
- RL environments = worlds (docs/slides/messages) + high-fidelity app clones (Salesforce, ServiceNow, Microsoft 365) via MCP/CLI/CUA + tasks with verifiers (rubrics/unit tests) . Mercur's talent network produced 2.5M expert hours in Q2, accelerating, to cover the economy's world/app/task distribution; humans remain essential for task/verifier creation because models can't reliably grade themselves (e.g., a legal environment built with Latham & Watkins) .
- Post-training on the 1,800-task Apex Agents set (~$500k compute) improved a legal-domain score from 4.7% to 26% and generalized to GDPval and Apex v1 ; GLM 5.2 and Kimik 3 are now on the leaderboard, enabling vertical frontier intelligence . Task pricing ranges $50–$10,000 (often ~$2,000), with frontier labs buying ~50k tasks/month; Mercur also sells off-the-shelf data built with hundreds of millions in investment .
- Synthetic trajectories are the core of RLVR, but humans are needed to measure beyond the frontier; future shifts include ultra-long-horizon tasks (100–1,000 hours) and virtual coworkers, since ~60–70% of real job tasks involve other people but only ~1% of evals measure social interaction . Mercur predicts dozens of app-layer companies will build their own frontier models in the next 12 months, citing Cursor as the template , and recommends custom exclusive data sets with partners like Mercur over in-house buildout .
- Base Power Company (Austin-based residential battery + retail electricity) raised a $1B Series D at a $12B pre / $13B post valuation, co-led by Ribbit Capital, Valor Equity Partners, Addition, and JPMorganChase's Strategic Investment Group; the round was fully subscribed within a day and multiples oversubscribed . Under three years old, it is the second most valuable energy startup (behind Helion) and the most valuable energy startup selling a product today . (Disclosure: Not Boring's Packy McCormick is an investor .)
- Team: CEO Zach Dell; co-founder/COO Justin Lopas (ex-SpaceX/Anduril, admits his manufacturing background was low-volume — a "blind spot" for high-volume production) ; Head of Hardware Dino Sasaridis (13 years at Tesla, led Powerwall 3 mechanical engineering) ; Head of Manufacturing Andy Ross (ex-Tesla Director of Battery Manufacturing Engineering, then GM; joined Mar 2025, built a 134-person team ~80% from his prior network) ; Head of Deployments Dana Paz (Anduril alum) ; CLO Zina Bash (Alito clerk; former Special Assistant to the President for Regulatory Reform) .
- Product/vertical integration: Base canceled its outsourced Gen 2 battery to accelerate fully in-house Gen 3 manufacturing — exiting supplier contracts for ~$1M in penalties versus the ~$15M cancellation cost originally feared . The resulting Base Core claims to be the highest-capacity residential battery on the market: 39.2 kWh vs 13.5 kWh for Tesla Powerwall 3, 20 kW inverter, -22°F operating floor, 50 ms transfer time, and "by far the cheapest residential battery on the market" . Vertical integration is framed as the core moat: faster feedback/iteration loops and a two-step install (the "Disco" disconnect) that cuts scarce electrician time .
- AI infrastructure pivot: Next move is pointing the "Capability Cannon" at AI — "bringing the power to the compute, and two is bringing the compute to the power" . Head of Software Jared Greene says the mission is fixing a grid that is increasingly for "compute," and that Base is starting "Solution Number Two": distributed compute placed where the grid has headroom — "a lot of sand" across the spiderweb — to bring "a lot of compute online while keeping the grid healthy" . Dell's stated ambition: "the world's largest energy company and perhaps the world's largest company" ; in late 2024 (at a $4B valuation) he said Base could become a $400B company .
- Traction: In July 2026, 6,000+ new members (up ~40% MoM) and a record 2,000+ installs, with >3 sign-ups per install — still supply-constrained . Deployed 585 MWh, targeting 1 GWh by year-end; July alone installed 2x the company's cumulative installs as of April 2025 . Demand mix: homebuilders (Lennar ~30% of 2025 installs; 77 communities by Mar 2026, >65% take rate), utility partnerships (5 deals totaling 220 MW; 58% of July installs in regulated territories), and direct customers . Retail plan Base Energy reached 14,000 members in 3 months; total membership 30,000+ . Chicagoland launched June 24, 2026; a Chicago Tribune feature drove a record 1,437 leads / 172 sign-ups in one day . Factory target: ~1,300 systems/week by end of 2026 (5x today), ~30x by 2028 .
- Policy/regulatory: Texas SB1252 (which Base's policy team helped pass) preempts local permitting rules for residential backup systems — a playbook Base is now pushing into a federal permitting bill . Illinois' Clean and Reliable Grid Affordability Act sets a 3 GW battery target by 2030 and gives residential batteries PJM capacity-market access . The FCC added foreign-made inverters to its Covered List citing remote-shutdown vulnerability — a tailwind for Base's US-made inverter .
- Investor sentiment: Ribbit published a "Power Letter" structuring the opportunity as Make/Move/Store/Sell applied to both electrons and intelligence — evidence of an emerging energy × AI infrastructure investment thesis . Key risk flagged by a Base investor: "If Base doesn't work outside of deregulated markets, if it can't break into utilities, it becomes a less good business" .
Frontier LLMs have stagnated — are "genuinely horrible" — at long-form technical and non-fiction writing even as they went from okay to superhuman at coding and mathematics; the most famous writing-capable models remain older releases such as OpenAI GPT 4.5 and Moonshot's Kimi K2, and bigger models are expected to fix small errors without transforming capability . In hands-on experience writing a post-training textbook, the author attributes the gap to a lack of verifiable rewards and training data: the RLVR approach he calls "truly magical" for reducing compounding errors in math and code does not transfer to prose, and token-spending harnesses (e.g., Claude Code-style setups) are not expected to have a multiplicative impact . He frames this as a fundamental limitation for "geniuses in a datacenter" autonomous science: models cannot yet organize and present established science or express their full knowledge in underspecified problems, so near-term LLM-driven science will be low-hanging fruit and cross-field connections, not revolutionary insight — even as Anthropic announced Claude making progress on the Riemann Hypothesis, he cautions that model coverage of scientific problems is narrower than many think .
- Model field notes: GPT 5.5 Pro found deep typos across the author's 200-300 page manuscript PDF; Claude models were the more useful editors (more taste, better mental model of the task); Claude Fable 5 composes a poem with only a minimal reasoning-token plan and no inference-time scaling — frontier models are not yet treating writing as a hard task .
- Practical economics and timeline: AI saved only 10-20% of the author's textbook effort, and he expects the best textbooks to remain human-crafted for another 2-5 years; AI is strong at adapting/formatting educational content but not creating it from scratch, and the highest-quality educational work remains scarce .
- Daydream Dental, an AI + human service that runs dental back offices (insurance claims, billing), pitches dentists on recovering 15–20% more revenue; founder is a Stanford grad/dropout who grew up working in his mother's dental practice; reported ~$10M revenue in year two .
- Qoves (QOVES.com) is a $150 consumer "looksmaxing" service (with a 1M-follower Instagram) that scores facial traits like cheek activation and averageness; a host who bought it questions whether it's actually AI or outsourced human critique — a cautionary example of AI-labeled beauty apps .
- Repeated investment theme: overlooked niche B2B businesses elsewhere — generator installation/servicing for hospitals/nursing homes (potential $100M–$1B), conveyor-belt tread, and dental billing AI — offer outsized returns with little smart competition; founders should avoid crowded spaces like programmer to-do apps .
Fireworks AI CEO/co-founder Lin Qiao positioned post-training as the core way startups can 'own your own intelligence' rather than rent off-the-shelf APIs, because apps are now trivially cloned from screenshots and baking product-specific data/taste into a model preserves differentiation .
Qiao laid out the post-training progression — prompt, RAG, supervised fine-tuning, preference/DPO, RL, distillation — with RAG for dynamic facts, SFT for output/structure, preference tuning for product taste, RL for weak spots, and distillation for cost/latency .
Traction examples on Fireworks: Cursor does deep mid-training to post-training and shipped Composer 2/2.5 models on par with frontier labs ; Doximity's clinic AI topped the Stanford/Harvard clinical safety benchmark ; Genspark's deep-research/slide-generation model slightly beats frontier quality at 5-10x lower cost ; Heidi's healthcare-tuned model outperforms the state-of-the-art closed model .
Market signals: 2025 was the year of coding, and the trend is moving to general-purpose and domain-specific AI coworker applications across legal, finance, marketing, recruiting, sales, and customer support ; incumbent CFOs are blocking AI feature launches over cost, making post-training a cost remedy ; Qiao expects millions of specialized models, one per use case .
Timing guidance: hit product-market fit first on frontier models, then post-train after PMF when product data is meaningful; post-training can cut serving cost 5-10x, enabling 5-10x traffic with the same budget and avoiding 'scaling into bankruptcy' . Reward engineering is similar to software engineering, so teams should start early .
Cautionary flags: reward hacking is real (a model asked to minimize compilation errors generated zero lines of code), and eval quality plus training-to-serving alignment (numerics/optimization) are key failure points .
Sandbar, the startup behind the Stream voice ring, raised a $23M Series A led by Jason and Kindred Ventures, bringing total funding to $36M; founder/CEO Mina Fami confirmed a $10M seed in early 2025 and a $3M pre-seed in early 2024.
Founder Mina Fami has a strong pedigree: MIT Media Lab neural-interface research, then a startup Control Labs (wrist gesture devices using deep learning) that Meta acquired in 2019 and whose tech became Meta's neural band; he met co-founder Kirak there. His thesis is 'self-extension'—hardware/software that augments the user rather than acting as a companion or coworker.
Stream is a push-to-talk private voice ring (touchpad, mic, haptic) for personal notes, real-time chat, and dictation across apps. It uses a custom whisper model for private input and a mix of partner and in-house models for transcription/actions; data is encrypted in transit and at rest with zero retention by processors.
Early traction: beta users and customers are majority outside tech, including working mothers and people with dyslexia; the product is designed for general use in motion rather than passive meeting recording.
Competitive context: many AI voice wearables have struggled (Humane sold to HP, Rabbit had software issues, Friend faced backlash); Sandbar differentiates with explicit human-in-the-loop, push-to-talk control, and social acceptability.
Hardware strategy: fully custom design with an experienced manufacturing partner (not an off-the-shelf reskin), a deliberate and capital-intensive path.
A new self-posted research paper, "When prediction itself rewards shortcut sensitivity" (arXiv:2604.21395) , argues that when a nuisance feature genuinely helps predict the training label, supervised learning has an objective-level incentive to keep using it . In a Gaussian population model with target y = wₛᵀs + ρwₙᵀn + ε, ρ ≠ 0 means the nuisance helps prediction , and the paper proves a lower bound — with a decoder of finite Lipschitz constant L, representation sensitivity satisfies D̃(φ,σ) ≥ σ²ρ²/L² — so that sensitivity does not vanish with more data or better optimization . This theorem is deliberately narrow (Gaussian setting, linear target, MSE) and is not a theorem about deep networks or adversarial training .
On a small ViT trained from scratch on CIFAR-10 , sensitivity magnitude and orientation decouple: PGD adversarial training cut the Jacobian Frobenius norm 34.1 → 3.0 but worsened the class-layout metric TDI (lower is better) 1.05 → 1.35, while PMH achieved TDI 0.87 with Jacobian 10.7 ; PMH perturbs inputs with isotropic Gaussian noise and penalizes ‖φ(x)−φ(x+δ)‖², capped as a share of the training objective . A small Jacobian norm thus says nothing about whether the remaining sensitivity points in useful directions; CKA and intrinsic dimension did not expose this magnitude/orientation dissociation .
Counterexample and caveats: on QM9, atomic positions carry real task information and suppressing sensitivity to position perturbations hurts performance — "Make the representation invariant" is not a general objective; the first question is invariant to what . Author-stated limits: the CIFAR model is deliberately small, headline numbers span 3 seeds (some secondary experiments single-seed), and TRADES/explicit Jacobian penalties are missing as baselines ; the post is self-authored and names no institution .
YC CEO Garry Tan joined a16z's Anish Acharya to discuss how AI is allowing small teams to do more with less, and expressed deep optimism about AI-native companies and SF's future . The podcast frames early-stage-relevant themes: solo founders, vibe coding, and "founders being 400x themselves" ; treating every business task as a markdown file ("skillifying") ; "tokenmaxxing" as a near-term way to work ; and the next computing paradigm centered on voice, memory, and "harness wars" . Tan's follow-up post doubles down on the "markdown skill-maxxing" framing .
Upfront VC Mark Suster shared and largely endorsed colleague @pzakin's essay "The agents will climb the ladder," which argues AI agents will progressively automate higher-order work and offers an investment playbook. Engineers are already lifted from coding to specs/planning, with 10x output reported and some teams dropping IDEs . Agents will likely climb the ladder too, eventually displacing humans from the creation chain — a frontier-lab onboarding reportedly told a hire "Welcome to the last two years of your career" . Applications can follow users up the ladder or shift to infrastructure: Cursor moved from copilot to multi-file edits to autonomous tasks to agent orchestration . Investment paths: build the primary interface for defining work strategy , build agent-ergonomic infrastructure , or build agents directly — but labs with huge capex have an economic imperative to capture agent-layer value, even cannibalizing API customers . As software costs fall (tokens cheaper than labor), enterprise build-vs-buy shifts toward DIY; code is rarely the moat, and durable third-party software needs scale advantages in generating useful tokens plus a domain where marginal intelligence gains matter — which most incumbents lack . The author's forecast: agents climb in every digital-work vertical; the great companies will own the largest compute pools solving the most important problems . Suster also flagged a related controversial piece, "the model eats the harness" by @JeremySF, saying he's convinced on much of the argument with a few things missing .
Garry Tan predicts a "wave of exceptional older AI-native founders" — experienced engineers aged 35–45 who have "been around the block, built a lot of engineering," complementing the usual young-founder pipeline . He points to Peter Steinberger as a model of this profile — someone who "knows what to build" — and says one such person can "outperform an entire department of any Mag 7" .
- Paul Graham: Demo Day investors remember at most ~5 things about a startup, so founders should decide those 5 key points in advance and make them the "vertebrae" of the pitch .
- Jason Freedman recalls that PG personally edited/polished/rewrote every startup's pitch in "vertebrae sessions" about a week before demo day — doing this for ~28 companies in S09 and ~100 in W12 .
- PG says the exercise also helped startups: trying to cook up super-ambitious future plans for investors often resulted in genuinely good new ideas .
During Paul Graham office hours, a founder from 2011 whose company is still going and doing well told the current YC batch that the biggest test is your own psychology — can you keep going despite all obstacles. Graham said the new founders started to grasp how hard startups are, "not just from what he said, but his face as he said it" .
Garry Tan released GBrain v0.45.6.0, adding 17 new "brain skills" hardened through his personal OpenClaw agent with hundreds of thousands of markdown files; he frames GBrain as a personal AGI and says it now works with Codex and Claude Code (repo: https://github.com/garrytan/gbrain). He advises running GBrain with Codex/Claude Code as a separate agent rather than in the main coding agent — a personal AI like ChatGPT/Claude with its own git repo for memory and custom skills .
Paul Graham says the current YC batch is the 47th, and he has spoken to nearly every startup in them (missing only a couple during covid), writing new talk notes each time . Asked how much standard startup advice still holds in the AI era, he says "so far almost all of it" — the variance in advice between different startups in one batch is larger than the difference between pre-AI and post-AI startups . He adds that AI may eventually turn starting a startup upside down but hasn't yet; the core remains "get growth by building what users need," with the biggest change being the many new things you can build .
Dalton Caldwell announced he led the Series A in Event Horizon Labs (EHL), a new quantitative research lab founded by @ocolegro (Owen) . EHL's research direction is automating the type of work Owen used to do at Citadel, with the thesis that this work can be done more effectively by self-improving AI vs humans . Caldwell calls Owen one of the most talented AI researchers in the YC community and notes several teams are pursuing this research direction . More: video interview with Owen ; EHL site: ehl.markets .
At Startup School 2026, Physical Intelligence cofounder Chelsea Finn said robotics is entering its 'GPT era' — moving from specialized models toward general-purpose systems that work across tasks, robots, and environments . She cited reinforcement learning pushing robot throughput up 2x and systems running autonomously for hours . The key unsolved problem is reliability: robots can fold laundry, make espresso, clean kitchens, and assemble things, but not yet do so reliably for long periods without human oversight . Talk covers reliability, RL, learning from failures, self-improvement, memory, general-purpose robots, one model beating specialists, and compositional generalization . Video: https://youtu.be/cRZNwgvcWUg.
Y Combinator's Garry Tan is 'deeply optimistic about AI-native companies and the future of SF,' saying AI lets 'small teams do more with less' and may extend Silicon Valley's power ; his own path ran through Microsoft, Palantir, and YC, and he frames the Valley as giving outsiders permission to build their way in . In the a16z podcast, he advocates 'tokenmaxxing' — loading 800K–1M tokens per request with agents like Hermes Agent or OpenClaw to 'live in 2028 today' — and describes reusable 'markdown file plus code plus tests' as 'an employee' that runs in cron and 'will do the job perfectly every single time' . Other covered themes for early-stage investors: solo founders/'vibe coding,' turning business tasks into markdown workflows, and a 'harness wars' race around voice/memory in the next computer .
