Y Combinator

2026-06-08 · open the original ↗

• Multiple founders credit Startup School with their decision to apply to YC. Several describe it as the pivotal event that gave them "the courage to apply," with one saying their company's entire trajectory would look different without having attended first.
• Seeing peers their own age build companies made founding feel achievable. Attendees point to hearing speakers like Fei-Fei Li and John Jumper the prior year as what "ignited the passion" to start a company, rather than any specific technical lesson taught.
• The event is framed as an unmatched concentration of ambitious, skilled people. Attendees describe it as selecting specifically for founders who are smart, ambitious and technically capable, with the shared room of similarly driven peers cited as what convinced several of them to commit to starting now rather than later.

2026-06-06 · open the original ↗

• Emergent reached $100M in annualized revenue nine months after launch. Mukund Jha, the company's co-founder and CEO, says the no-code app-builder lets anyone ship real, deployed software — including a working backend and database — just by chatting with an agent, unlike most competing tools he says only produce front-end "demoware."
• Purposeless tinkering after his last company's collapse led directly to the idea. Jha says he spent up to twelve hours a day experimenting with new AI models with no business objective following the 2023 shutdown of his earlier startup Dunzo, and that period of pure curiosity generated the technical insight that coding itself was about to be disrupted.
• Twelve VCs rejected the pitch to "automate software engineering." Jha says the team deliberately built toward where AI models were extrapolated to be in six months rather than where they currently stood — for instance skipping a JSON-output-parsing problem roughly 20 to 30 other YC companies were separately trying to solve, betting correctly that the next model generation would fix it.
• The team chased the world's hardest coding benchmark before finding a product. Jha says he picked SWE-bench mainly to distract a frustrated team from pivoting every week, but becoming the world's top-ranked system on it with a four-person team became the technical foundation — including parallelized test-time compute and agent memory — that Emergent's product was later built on.

2026-06-05 · open the original ↗

• Paxel analyzes a builder's own coding transcripts to profile how they work. It scores users across five dimensions — steering, execution, engineering, product instinct and planning — surfacing patterns like their most-used prompt or how many agents they run in parallel, framed as "a mirror, not a grade."
• All analysis runs locally, so a user's code never leaves their machine. The tool is described as a single command that takes 15 to 30 minutes to run before emailing the resulting builder profile, positioned as a way to surface skilled builders that a traditional resume would never reveal.
• YC invites Startup School applicants to submit their Paxel results. The team frames this as a way to see "how you build, not just what you wrote," telling applicants who already submitted to go back and add their token, and stating it can only help, never hurt, an application.

2026-06-04 · open the original ↗

• He no longer writes code, only orchestrates agents. Charlie Holtz, co-founder of Conductor, says he manages multiple parallel coding-agent workspaces by voice, kicking off tasks with a gooseneck microphone and reviewing PRs rather than editing files himself; a manual-edit mode exists but is called "caveman mode" for how rarely it's used.
• Deliberate "slop-free zones" stop AI from architecting the product. Conductor keeps parts of the codebase and docs, plus UI and interface decisions, off-limits to AI contribution and human-reviewed line by line, because he argues unattended AI code can spiral into a vicious cycle where bad code begets more bad code.
• He spent $22,000 on tokens in a single month building Conductor, in July 2025, on a prior model generation, and now advocates "token maxing" — fast mode, high reasoning effort — while deliberately keeping lines of code minimal, since an established codebase can spiral out of control if line growth isn't controlled.
• Code is becoming disposable; the prompt is the real asset. He argues code is now "sawdust" left over from describing what you want, so when new models ship you can simply rerun the same prompts to regenerate better code, likening this "malleable software" to modding a video game around a fixed core structure.

2026-06-03 · open the original ↗

• The best AI-services markets share four specific traits. The speaker, presenting a YC playbook, argues founders should target work that is already outsourced (low trust), breakable into automatable steps (low judgment at the task level), genuinely hard (a high intelligence threshold), and regulated — since regulation raises the bar and the moat rather than just adding friction; examples cited include tax, audit, insurance, mortgages and healthcare.
• Founding teams need domain fluency, model fluency, and operational rigor — the product is the operation. He points to General Legal Team, an AI-native law firm YC backed, whose founders combine law-firm experience at Cooley and Fenwick with technical leadership from CaseText and have built shift-based staffing to cut cycle times, calling variance in output the "existential problem" that causes customers to fire a vendor faster than slowness or price ever would.
• New AI-services founders should deliberately cap their first pilot customers. He warns of an "early demand trap" where easy early sign-ups overwhelm delivery capacity before the product can scale, and advises pricing on outcomes or per-unit (per return, per claim, per loan) rather than cost-plus or straight-line undercutting, since services compete against the cost of labor, not other software.
• The core financial bet is "AI operating leverage" pushing services margins toward software's. He argues traditional services firms cap out near 30% margins while AI-driven ones can approach 50%+ as the product lowers cost of goods sold (models, hosting, humans), on markets two to three times bigger than software's; he also argues buying an existing services business to bolt AI onto rarely works, since legacy hiring and performance expectations don't change with it.

2026-05-29 · open the original ↗

• YC's Harj rejected their EdTech idea, not their candidacy, and bet on the founders instead. Varun, GigaML's co-founder, says Harj told them "you guys are really good engineers, pick something else" without asking about their idea or traction, pushing them to abandon EdTech and pivot within a month of starting the batch based on their LLM research background rather than a specific pitch.
• A GitHub usage pattern, not a market thesis, drove the pivot into customer support. After their initial fine-tuning product proved a bad market — good only for cutting inference cost or serving security-sensitive buyers that are hard for engineers to sell to — Varun says they noticed customer support and coding were the two use cases growing fastest among their fine-tuning customers, and Zepto became their first customer once they switched.
• An eight-person team beat a roughly 400-person, well-funded rival for DoorDash's contract. Varun credits a YC-network "unfair advantage" (an introduction from Garry Tan) and a clean three-month pilot with no downtime, and says winning DoorDash proved to him that building a superior product beats building a sales team — a lesson he calls an "arbitrage" most competitors were missing.
• Coding agents let GigaML run with roughly one-seventh the engineers it would otherwise need. Varun estimates the company would need six to seven times its current engineering headcount without tools like Claude Code, and says the deeper benefit is avoiding context-switching and context transfer between people, not just cost; the company's interview process still has candidates write code with AI, then edit it without AI, to confirm they understand it.

2026-05-28 · open the original ↗

• Inference speed will become a capability ceiling, not just a cost lever. Presenting his "speculative speculative decoding" work, Stanford grad student Tanishk argues that once performance scales with amount of thinking, tokens-per-second is peak intelligence delivered; his method parallelizes normally-sequential drafting and verification by predicting verification outcomes about 80-90% of the time, winning on both latency and throughput sampling Llama 3 70B on 4 H100s.
• Factoring the action proposal from the dynamics model lets a robot policy adapt at runtime. Google DeepMind's Stannis presents Diffusion Model Predictive Control, where multi-step diffusion models for action proposals and dynamics reduce compounding errors and simplify planning; he shows a simulated walker with a broken ankle recovers most of its performance just by re-adapting the dynamics model, without retraining the action proposal.
• A single new regularizer lets a world model avoid representational collapse with one hyperparameter. Isaac Ward, from Yann LeCun's group (whose $1.03 billion raise centers on world models), presents "LeJEPA," which adds a "SIGReg" term enforcing Gaussian-distributed latent embeddings; he reports it runs roughly 50 times faster than competing world models on a single GPU under 24GB of memory with just 15 million parameters, and can flag its own prediction errors when conditions shift.
• Ensembling and distillation, not bigger single models, are the efficient path when data — not compute — is the constraint. Presenting a paper co-led with Suhas and Percy Liang plus another collaborator, Ku shows that regularized and ensembled pre-training recipes follow clean power laws with measurable best-possible-loss asymptotes; a joint recipe yields roughly a 5x data-efficiency win, and distilling an eight-model, 2.4-billion-parameter ensemble into one 300-million-parameter model retains about 83% of the gain.

2026-05-27 · open the original ↗

• A single shared database, not better tools, was YC's real unlock for internal agents. Partner Pete Koomen says YC runs on its own software sitting on one Postgres database covering every company, founder and financial record, so an agent with read-only SQL access could answer arbitrary questions; he argues this collapsed the friction of cross-team requests enough to trigger a Jevons-paradox-style explosion in the number and complexity of questions people bothered to ask.
• A tool registry grew from about 20 tools to over 350 as teams kept adding their own. Koomen describes YC's internal harness evolving from a simple agent loop into a shared registry that any team can extend, and says the key move was building a "resolver" — an agents.md-style entry point linking to markdown skills — so agents could discover and reuse capabilities instead of duplicating them, checked for redundancy with what Garry Tan calls a "dry and MECE" meta-skill.
• YC deliberately broadcasts every internal agent conversation to Slack by default. Koomen says this transparency, chosen despite internal debate, let employees learn techniques by watching others use agents and created a social check that made it safer to give agents broad access to sensitive context — a trade-off he argues requires an "egalitarian, trust by default" culture that works best in small, aligned organizations.
• A single partner's handwritten skill for writing "two-sentence descriptions" became better than any individual partner at YC. After other partners fed a group office-hours transcript back to the agent to refine the skill, Koomen says it now writes better two-sentence founder pitches than he does personally, illustrating what he and Tan call building "superintelligence inside a company": composing many individually improved skills into a shared organizational capability.

2026-05-22 · open the original ↗

• Shareholder primacy is a recent legal invention, not a law anyone ever passed. Eric Ries, author of the forthcoming "The Uncorruptable," traces it to a small group of academics and judges in the 1960s-70s who reinterpreted "any lawful activity" charters as requiring shareholder-value maximization — a norm he says has never been subject to legislation or referendum, yet is treated as binding by courts, as when Twitter's board was compelled to sue to complete Elon Musk's takeover.
• Founders keep losing companies they built even while succeeding by every normal measure. Ries cites Twilio's Jeff Lawson, ousted 199 days after his dual-class control sunset expired despite revenue growth since IPO, and Sol Price, locked out of FedMart in 1975 by a new board after 20 years building it; FedMart went bankrupt within seven years of adopting standard shareholder-value practices, while Price's next venture became Costco.
• Independent board directors don't solve investor control — they just make it look balanced. Ries argues independent directors have no equity stake in a company's mission but do have a financial incentive to be seen as pro-investor, since investors recommend them for future board seats, making the classic two-VC/two-founder/one-independent board effectively investor-controlled; he cites data that companies using a separate nonprofit-trustee structure instead are six times more likely to survive to year fifty.
• A nonprofit trustee board can preserve value the market would have destroyed. Ries recounts Novo Nordisk's foundation trustees blocking a lucrative 2000s merger because it wasn't necessary for the company's survival — the rival company that merged instead had its R&D cancelled after being acquired by Merck — allowing Novo Nordisk to keep funding an unproven GLP-1 program for over a decade to a later $600 billion valuation; he points to Anthropic's similar long-term benefit trust as enabling its refusal of a $200 million contract.

2026-05-21 · open the original ↗

• OpenAI is investing $2 million in tokens, not cash, into every spring and summer 2026 YC company. Garry Tan explains the money comes via an uncapped SAFE, effectively pricing the investment at the company's future Series A valuation but committing it at the earliest stage, and can be spent on Codex or the OpenAI API.
• The deal targets founders "token maxing" through software-factory patterns. Tan describes dispatching many agents to burn large volumes of tokens as a way to reach what would have been 18 months of engineering and Series-A-level product polish in a fraction of the time, and says YC observed its best current companies already spending heavily on tokens this early in their life cycle.
• YC extended a special application window since the summer batch deadline had already passed. Tan says the offer, developed together with Sam Altman, is being run as an experiment for just this spring and summer batch, with applications due by Memorial Day weekend (May 25) and decisions promised by June 5.

2026-05-19 · open the original ↗

• Zepto's founding pivot came from data, not strategy: one dark-store neighborhood outperformed the rest 3-4x. Aadit Palicha says he and Kaivalya Vohra started at 17 delivering groceries via a WhatsApp group in Mumbai during COVID, under the name Kirana Cart, but only switched to running their own mini-warehouses — the first literally inside a co-founder's apartment — after seeing that neighborhood's order volume dwarf areas still served by doorstep delivery from existing stores.
• They designed backward from an extreme ideal experience rather than forward from what was operationally feasible. Palicha credits a discussion of Airbnb's "seven-star" thinking for the approach behind 10-minute delivery, and argues that nailing customer experience — not cost-cutting — created the demand and throughput per warehouse that then made costs low, concluding "customer delight is where financial value starts."
• He delayed dropping out of Stanford for about eight to nine months until reaching early product-market-fit signals. Palicha says he waited until the business hit roughly 10,000 orders a day and a run rate of 60-70 crore, plus secured a term sheet, before committing full-time — advising founders to have proof a business is working before taking that kind of risk, not before.
• AI operations work has let Zepto's supply chain and corporate teams run with far fewer people. Palicha says machine-learning forecasting now runs dark-store replenishment for millions of units a day without humans in the loop, replacing forecasts that used to take teams days, while GenAI-driven ad-keyword tools boosted advertising revenue and internal automation let the company cut nearly all SaaS and outsourced-service spending with a corporate team of about 500 engineers plus 150 in data, product and design.

2026-05-16 · open the original ↗

• YC says the India event drew over 25,000 applications. A speaker tells the crowd that India's AI ecosystem will be defined by the people in the room, not by the prior generation.

2026-05-13 · open the original ↗

• Every field concentrates in one place, and the historical answer is always to go there. Paul Graham compares Silicon Valley today to Paris for painting in 1870 or Göttingen for math in 1900, arguing the talent pool at a center improves in two dimensions — better people and more of them — which he says produces disproportionately valuable serendipitous meetings that he cannot fully explain but has observed repeatedly in biographies of accomplished people.
• Silicon Valley investors decide faster because competition punishes hesitation, and it doesn't hurt their returns. Graham argues the more confident an investor is about an opportunity, the less time they can afford to wait, illustrated by a Boston VC firm that faxed Dropbox a blank-valuation term sheet only after learning Sequoia was interested — too late, since Dropbox went with Sequoia and became the first YC company to go public, in 2018.
• Comparing yourself to visible "big fish" resets ambition from impossible to merely hard. Graham says moving to a big pond lets founders measure themselves against known peers like Brian Chesky or Sam Altman and conclude they aren't a different species, just people who worked exceptionally hard — turning a fogged-in, seemingly unreachable summit into "a high but definite threshold."
• Founders should go to Silicon Valley, ideally via YC, and then return home — even though data shows returners do worse on paper. Graham argues going and coming back is the best way to help a hub like Stockholm, bringing back money, imported culture and a stronger startup, but acknowledges YC data shows founders who return home are only about half as likely to become unicorns as those who stay — a gap he attributes partly to selection bias and to Bay Area startups raising at higher valuations, not necessarily performing better.

2026-05-08 · open the original ↗

• Garry Tan rebuilt the same blog platform three times, and the cost collapsed by roughly 20,000x. He says the 2008 version (Posterous) took $4 million, six or seven people and about a year and a half; a 2013 rewrite took roughly $100,000, two people and three months; the 2026 version, built with a Claude Code Max subscription, cost about $200 and five days while adding full agentic retrieval and research capability the earlier versions never had.
• "Boiling the ocean" — paying for exhaustive completeness — is now cheap enough to be standard practice. Tan argues that because an agent can cross-reference twenty sources as easily as one, builders should stop economizing on research or code depth the way a human effort-budget would force them to, calling this "token maxing" and arguing it will spread from coding into every kind of knowledge work.
• A "thin harness, fat skills" philosophy puts judgment in markdown, not code. Tan describes building G Stack around markdown skills — a "CEO" review for ambition, a plan-and-engineering review, a developer review, and browser-based QA — while keeping the underlying agent harness minimal, arguing that anything requiring human-style judgment belongs in plain-language markdown rather than brittle deterministic code, and that skipping 80-90% test coverage produces output "10x worse than human-written code."
• Measuring logical lines of code, Tan found he was producing about 400 times his 2013 output, directing roughly 15 agents at once, even though the code he personally typed dropped about 70% from his 2013 baseline — a distinction he says got lost when his "100x" claim drew online criticism. He frames the larger stakes as builders needing to control their own prompts and tools rather than depend on centrally controlled AI products, calling it "the defining question" of the moment.

2026-05-08 · open the original ↗

• YC is hosting a one-day Startup School Paris on June 29th. Builders across France and Europe will hear from founders of companies including DataDog, Supabase, and Amie, plus get direct access to YC partners and alumni.

2026-05-06 · open the original ↗

• Regulatory friction that nearly killed Razorpay's launch became its deepest moat. Harshil Mathur says it took a full year after joining YC just to complete their first live transaction, since payment licensing required certifications and approvals before any transaction could occur; he argues this hurdle is actually fair because every later competitor must clear the same year-long process, which keeps the field of entrants small in a way ordinary tech products can't replicate.
• A banking-partner crisis two weeks after demo day became a lesson in trust that shaped a permanent AI policy. When Razorpay's enabling bank abruptly cut off around 50 live merchants over one complaint, Mathur says the team personally called every affected customer, absorbing abuse rather than going silent, and regained a live partner within days; he cites this as why Razorpay still refuses to let AI replace customer support, even while using it elsewhere, because he sees support as a trust channel rather than just a problem-solving function.
• Razorpay deliberately stayed capital-inefficient by choice, banking Series A funds instead of burning them. After raising roughly $11 million, Mathur says the company kept monthly burn under $200,000 and earned more in bank interest than it spent, angering an investor who expected it to burn faster; he argues B2B economics are inherently "logical" — a business pays only for value delivered today, unlike consumer models that spend to acquire and monetize later.
• A calculated early bet on India's then-unproven UPI system, made because the company had nothing to lose, created a six-month competitive window. Mathur says Razorpay went live on UPI months before India's largest banks did, so when demonetization hit in November 2016 and demand surged, it was the only payment gateway ready — winning clients like Zomato, Swiggy and BookMyShow before rivals caught up; he extends this into a broader claim that AI is now compressing capital and infrastructure moats industry-wide, leaving execution speed and the willingness to reinvent proactively, rather than wait to "respond," as the only differentiator left.

2026-05-04 · open the original ↗

• Current AI chips are poorly suited to how agents actually run. YC says this causes GPUs to hit only 30–40% peak utilization on agentic workloads, because agents loop, branch, and backtrack across dozens of steps, bouncing between memory-bound model calls, IO-bound tool use, and CPU-bound orchestration — a fundamentally different hardware problem than simple prompt-in, response-out inference.
• Purpose-built silicon for the agent loop itself is still missing. YC points to Nvidia's reported $20 billion acquisition of Groq and Google's inference-specific TPU v7 as evidence the shift toward specialized hardware is underway, but argues the real advantage is the compiler, not the chip — fast context switching between models, native speculative decoding, and memory built for KV caches that persist across an entire execution graph.

2026-05-04 · open the original ↗

• Scientific discovery is shifting from AI copilots to closed discovery loops. YC argues this is already happening in drug discovery, material science, and protein engineering, where models propose candidate molecules, automated labs synthesize and test them, and results feed back to iteratively improve the next candidate; it predicts the most impactful companies won't just sell research copilots but will be full AI-native discovery engines.

2026-05-04 · open the original ↗

• The best AI-native companies make their whole operation queryable by AI. YC argues this turns decision-making into a closed loop — the system monitors what's happening, compares it to what should be happening, and adjusts — citing teams that cut sprint time in half and shipped 10x as much after recording every meeting and tracking every ticket and customer interaction.
• Building this today requires brutal custom integration work across tools. YC points to Slack, Linear, GitHub, Notion, and call recordings as the kind of systems currently stitched together with custom glue code, with no single product connecting it all into one AI layer; it wants to back the connective layer that makes a company legible to AI by default, rather than another one-off patchwork.

2026-05-03 · open the original ↗

• AI has collapsed the cost of producing software by 100x or more. YC argues this erases the code-volume moat that protected legacy SaaS incumbents, framing the recent hit to software market caps as bad for incumbents but good for startups, since decades-old codebases once considered untouchable are now vulnerable to cloning, bundling, or open-source replacement.
• YC is pushing founders toward bigger targets than typical product-management tools. It names chip design software, ERPs, industrial control systems, and supply chain management as the next frontier, framing it as a repeat of the shift from on-premise to cloud: the next generation of great software companies will replace legacy SaaS with AI-native software.

2026-05-03 · open the original ↗

• Star Cloud's Adi Oltean wants YC to fund lunar industrial capabilities. He describes starting with extracting raw materials like silicon, aluminum, iron, and titanium from lunar regolith via electrolysis, then 3D printing complex structures from the molten material — which he says could be more efficient than on Earth due to the lack of structural supports needed.

2026-05-02 · open the original ↗

• Every major software category needs to be rebuilt for AI agents. YC argues agents need machine-readable interfaces like APIs, MCPs, and CLIs, plus thorough documentation to discover and use tools without a human in the loop, replacing the visual forms and dashboards built for people — and predicts this rebuild will come from agent-first startups rather than incumbents bolting on agent support.

2026-05-01 · open the original ↗

• US hardware moves far slower than China's because of iteration speed. YC contrasts Shenzhen, where a team can go from design to a physical part in a day, with the US, where the same loop takes weeks, attributing China's edge to dense supplier networks and tight coordination between design and production.
• YC wants to fund startups that make US hardware iteration dramatically faster. It points to a startup already building faster turnaround from design to manufactured mechanical part as an early example of the kind of missing infrastructure it's looking for.

2026-05-01 · open the original ↗

• Transformers trade recursive reasoning for parallel training speed. Chopard argues that unlike RNNs, which iterate on a compressed hidden state across time steps, an LLM computes in one feedforward pass with no compression in the time direction — every decode re-processes the full context. RNNs paid for their recursion with backprop-through-time, which caused vanishing and exploding gradients as models or sequences grew.
• Some problems are provably incompressible, unsolvable in one forward pass. He cites comparison sorting's known n·log(n)-step lower bound, plus Sudoku, mazes and rolling sums, arguing a transformer with fewer layers than the required steps cannot solve them. Chain-of-thought and tool-calling can simulate a Turing machine at test time, but only within human-demonstrated traces — an LLM taught only bubble sort, he says, could not discover merge sort on its own.
• A tiny recursive model outperformed giant pretrained LLMs on ARC-AGI. He credits the Hierarchical Reasoning Model — a 27-million-parameter model with no pretraining, trained on roughly 1,000 ARC-AGI tasks — with scoring about 70% on ARC-1 versus OpenAI's o3 scoring zero at the time. The follow-up Tiny Recursive Model, by a researcher he calls Alexia, shrank it to 7 million parameters by collapsing HRM's two networks into one, and pushed the score to 87%.
• The outer refinement loop, not the two-network design, drives the gain. Chopard says a researcher at Francois Chollet's company ran scaling ablations showing this repeated outer loop is the main source of HRM and TRM's performance, which is why TRM kept and simplified around it while discarding most of the rest. He expects the next leap to come from combining this recursion with giant pretrained LLMs rather than treating either as a substitute for the other, since today's recursive models are trained per task, not general-purpose.

2026-04-30 · open the original ↗

• AI has erased the barriers to selling to the world's largest companies. YC argues enterprise leaders now actively seek teams that can apply AI to their problems, coding agents let a two-to-three person team ship nuanced products to large companies in months rather than years, and those leaders now understand what to build in-house versus outsource — making it common for a YC startup's first customer to be one of the largest companies in the world. This, YC says, kills the old model of startups operating in stealth for years to reach feature parity with an incumbent before launching.

2026-04-30 · open the original ↗

• Semiconductor supply chains are still run on spreadsheets and calls. YC notes a single advanced AI chip takes 1,400 manufacturing steps across a dozen countries over five months, and cites 2021, when a $300 chip held up a $50,000 car and left $200 billion in vehicles unbuilt, as proof that companies see their direct suppliers but have no visibility into second- and third-tier ones.
• The bottleneck has only gotten worse. YC says Nvidia has locked up over 60% of TSMC's advanced packaging capacity, high-bandwidth memory is booked through 2026, export controls change quarterly, and the CHIPS Act is standing up new fabs in Arizona, Texas, Ohio, and New York that each need a supply chain built from scratch.
• This tooling doesn't exist yet, and building it needs deep technical knowledge. YC says real-time allocation tracking, multi-tier risk monitoring, and export compliance all require understanding wafer allocation and packaging constraints at a deep level — which is why it sees this as a startup opportunity rather than a feature bolted onto SAP.

2026-04-29 · open the original ↗

• Coding agents now let ordinary users become their own "forward-deployed engineers." YC notes that before AI, nearly all software shipped one interface for everyone — Netflix looks the same for every user, just with different imagery — with deep per-customer customization limited to enterprise software built by human forward-deployed engineers, and argues users can now radically customize interfaces themselves.
• This implies rethinking how software gets delivered. YC raises open questions like whether companies will need to ship source code instead of packaged binaries, and whether users' agents should be able to modify only front-end elements or also middleware on the fly.

2026-04-29 · open the original ↗

• YC is launching a Paper Club to keep builders and researchers current on frontier AI. Framed around how hard it is even for a PhD student to keep up, the group will meet every few weeks at YC's Mountain View office to present and discuss papers, followed by an off-the-record dinner; the first cohort caps under 100 people and meets May 20.

2026-04-29 · open the original ↗

• Hassabis expects AGI soon, but flags one or two open unknowns. His personal timeline is around 2030; he says continual learning, long-term reasoning and some aspects of memory remain unsolved, and puts it at "50/50" whether closing those gaps needs one or two genuinely new ideas or just incremental scaling of what already works — DeepMind is pursuing both bets at once.
• Today's reasoning models are only "fairly brute force," producing jagged intelligence. Watching Gemini play chess, Hassabis says its own thinking trace will identify a move as a blunder and then play it anyway — evidence, he argues, that models lack introspection into their own reasoning, which is why the same system can solve IMO gold-medal problems yet still make elementary math errors.
• Agents haven't truly arrived until a vibe-coded game tops the charts. Hassabis says no AI-built game has hit the App Store charts the way a studio-made hit would, and expects that to change in the next 6 to 12 months; he predicts individual builders working at roughly 1000x their prior output will appear before agents operate fully autonomously.
• The AlphaFold pattern: combinatorial search, a clear objective, plenty of data. Hassabis says that formula — a massive combinatorial search space, a clear objective function to hill-climb, and enough data or simulation to generate synthetic training data — drove both AlphaGo and AlphaFold, and advises founders to pair deep AI expertise with a hard science domain like materials or medicine as one of the few bets defensible against the next foundation-model update.

2026-04-28 · open the original ↗

• Space-optimized inference chips are an emerging startup opportunity. Philip Johnston, co-founder and CEO of Starcloud (building data centers in space), argues reusable rockets from SpaceX and Stoke Space will drive a huge increase in orbital compute capacity, creating demand for chips tuned for mass, thermal and radiation constraints; he's looking to hear from chip designers with experience at SpaceX or Nvidia.

2026-04-28 · open the original ↗

• A drone swarm reportedly destroyed an AWS data center unopposed. He says a swarm of cheap Iranian drones took out the facility last month with nobody stopping it, and warns that a coordinated swarm of a thousand drones — far larger than the single hobby drones most counter-drone systems are built to handle — would be even harder to stop.
• The cost math today overwhelmingly favors attackers, not defenders. He points to a $3 million Patriot missile intercepting a $500 FPV drone as the current exchange rate, and argues restoring that cost advantage to the defense side is the real problem to solve.
• YC wants to fund the counter-swarm stack as software, not weapons. He lists interceptors that neutralize dozens of drones at once, sensor-fusion software giving a single real-time picture, and non-kinetic tools like rotor-fouling aerosols and entangling streamers — arguing that with radio jamming becoming obsolete, the winning companies will resemble Cloudflare more than Raytheon.

2026-04-28 · open the original ↗

• AI automation of companies is now blocked by scattered internal knowledge, not model quality. The speaker argues every business needs a "company brain" — a system that extracts tacit know-how from emails, Slack, support tickets and databases into a living, current "skills file" so AI agents can execute tasks like refunds or pricing exceptions safely and consistently.

2026-04-27 · open the original ↗

• Falling diagnostic and gene-therapy costs, paired with agent harnesses, will enable personalized medicine at scale. The speaker says genome-sequencing costs have fallen faster than Moore's Law and that N-of-1 mRNA-based therapies plus greater FDA openness are emerging, arguing agents like Claude Code can combine genomic, EHR and wearable data into individualized treatment suggestions.

2026-04-27 · open the original ↗

• AI-native companies that sell the finished service, not a software tool, are the next wave. The speaker contrasts 2023–2025 "AI co-pilot" startups with a new category that performs the work itself — citing insurance brokerage, accounting, tax, audit and healthcare administration — arguing total services spend dwarfs software spend and is easier to disrupt because much of it is already outsourced.

2026-04-27 · open the original ↗

• Farmers are trapped in a worsening chemical treadmill. The narrator argues pesticide resistance keeps evolving, the pipeline for new chemicals has gotten slower and more expensive, and farmers must keep spraying more for diminishing results even as costs rise and margins fall, because failing to spray risks crop loss and food-price shocks.
• Three technologies converging at once make the problem newly solvable. He points to cheap sensors and cameras giving AI real-time weed/pest identification, robotics precise enough to treat individual plants instead of whole fields, and biological alternatives (microbes, peptides, RNA-based tools, engineered self-defending plants) capable of replacing whole classes of synthetic chemicals.
• This is framed as a generational investment opportunity for YC. The narrator predicts adoption will be "explosive" rather than slow because a company that cuts pesticide use by 90% while increasing yields would be attractive in one of the world's largest markets, and closes with an open call for founders building in this space.

2026-04-25 · open the original ↗

• Replit deliberately targets tech-adjacent creators, not classic developers. Amjad Masad, Replit's CEO, says the users getting the most value are product managers, designers and entrepreneurs who don't want to touch environment setup or deployment, and in 2023 the company made it an explicit goal to stop building for traditional developers in favor of what he calls "AI-native developers."
• Enterprise clients are reporting an order-of-magnitude jump in idea throughput. Masad cites Whoop as an example, saying the company used to generate around 100 product ideas but could only afford to try five of them, and can now try roughly 50 because product and design staff, not just engineers, can build working software themselves.
• Agent 4's redesign targets the "boredom" of watching an agent work. Masad says the release adds parallel agents, an asynchronous design canvas so users can keep working while code builds in the background, and multiplayer sessions where each collaborator forks their own virtual machine, all aimed at autonomy running for hours at a stretch.
• He predicts a "post-prompting" future built on just builders and salespeople. Masad expects agents will soon take high-level goals rather than detailed prompts, with salespeople surviving as evangelists and educators because customers still trust humans, and builders becoming generalists who direct agents — citing an internal Replit team that improved support-ticket prioritization and built an HR onboarding tool without a defined mandate.

2026-04-24 · open the original ↗

• AI-native companies should run as "closed loops," not open ones. Diana, a YC partner, argues every process should feed its outcomes back into an AI layer with full context — meetings, dashboards, Slack, tickets — rather than relying on manual status updates; she cites YC companies giving an agent access to Linear, Slack and customer feedback to generate sprint plans, cutting sprint time in half and getting close to 10x more done.
• "AI software factories" replace code with specs and tests agents iterate against. Diana describes this as the next evolution of test-driven development, where humans define success criteria and judge output while agents write and iterate on the implementation, pointing to StrongDM's engineering team as having pushed this far enough that their repos contain no hand-written code.
• Management hierarchy stops making sense in a fully AI-legible company. Diana relays Jack Dorsey's view at Block that companies must be rebuilt around an intelligence layer rather than human routing, replacing traditional management with three roles: the individual contributor who builds, the directly responsible individual accountable for one outcome, and a founder who leads by using the tools directly.
• Startups have a structural edge over incumbents in going AI-native. Diana argues early-stage founders can design workflows and culture around AI from day one, unlike established companies that must unwind existing systems and processes — citing Mutiny's use of a separate internal skunkworks team as one way larger companies try to work around that constraint.

2026-04-23 · open the original ↗

• The bottleneck in AI coding is thin scaffolding, not model intelligence. Garry Tan, YC's president and CEO, argues Claude Code's models are already capable enough that guessing at an unfamiliar codebase — not weak reasoning — produces plausible-looking code that silently breaks, which is why he built G-Stack around a "thin harness, fat skills" approach that has passed 70,000 GitHub stars.
• The "Office Hours" skill reframes an idea the way YC partners do. Demoing a tax-document app idea, Tan shows it asking for evidence anyone wants the product, questioning why TurboTax or Plaid don't already solve it, and reframing a simple document aggregator into a wedge toward a CPA-marketplace lead-gen business.
• Adversarial review automatically caught and fixed most gaps in a design doc. Tan says the design document survived two rounds of multi-step adversarial review that caught 16 issues, most auto-fixed, raising its self-assessed score from 6/10 to 8/10 before he approved it.
• Tan runs parallel Claude Code sessions and ships dozens of PRs daily. He says he runs 10 to 15 sessions at once and ships up to 50 PRs in a day, after building a custom CLI wrapper around Playwright and Chromium for browser-based QA when the Claude-in-Chrome MCP tool proved too slow and unreliable — a setup he describes as reaching "level seven" of an eight-level software factory.

2026-04-22 · open the original ↗

• Stripe's product suite outgrew its old site's story. Katie Dill, Stripe's head of design, says the 2020 homepage still looked current but no longer conveyed a business that had expanded from payments into billing, tax, revenue recognition, usage-based AI billing and stablecoins — not the design aging poorly — a shift she notes drives use by over 78% of the Forbes AI 50.
• A live GDP counter was a deliberate, hard-to-copy trust signal. Dill says the team added a live counter showing the share of global GDP that runs through Stripe to demonstrate reliability at scale to customers ranging from micro-entrepreneurs to large enterprises, choosing that specific metric because few other companies could credibly claim a comparable figure.
• User research killed an accordion-style layout because people simply don't click tabs. Dill says testing showed the tap-to-expand format required too much effort for visitors in "browse mode," which pushed the team toward the more visual "Bento" grid with pop-up modals that shows products directly rather than requiring interaction to reveal them.
• AI sped up prototyping but didn't replace design judgment. Dill says the team could review roughly 20 visual directions in the time it used to take to review two, using a custom internal tool to tune the homepage's gradient-wave visuals, but still chose to push the launch past December because a set of card animations didn't yet feel smooth enough.

2026-04-16 · open the original ↗

• One model trained across many robot bodies beat single-robot specialists. Qiong (Quan) Wang, co-founder of Physical Intelligence, says their Open X Embodiment experiment trained a generalist policy on data from 10 different robot platforms and found it performed 50% better than ten specialist policies each optimized for a single platform.
• Physical Intelligence's robots run on a model hosted in the cloud. Wang says nearly all their demos, including laundry folding and mobile navigation, query a real cloud data center over an API within the control loop, made workable by an algorithmic technique he calls "real-time chunking" that precomputes overlapping action sequences so network latency doesn't stall the robot.
• Two YC-backed deployments show the mixed-autonomy approach already running real commercial work. Wang describes Weave's robot folding varied, unseen laundry items in a real laundromat, and Ultra's system packing e-commerce orders into shipping pouches inside a live warehouse for a full day with only minimal human intervention — not lab demos, but real customer orders being fulfilled.
• Wang predicts a "Cambrian explosion" of vertical robotics startups. He argues cheaper hardware, models reactive enough to compensate for imprecise motion, and a path to break-even economics via mixed human-robot operation let a small, scrappy team build a vertical robotics company without the full vertical integration the field used to require — and Physical Intelligence open-sourced its Pi Zero and Pi 0.5 model weights, identical to the versions used internally, to accelerate it.

2026-04-06 · open the original ↗

• A synthetic-DNA spike-in turns a biology problem into a math problem. Co-founders Ozan and David say BillionToOne adds known synthetic DNA to a blood sample before PCR amplification, which lets them measure exactly how much distortion and error the amplification process introduces at each genomic location and mathematically subtract it — the technique behind a prenatal test now processing over 600,000 tests a year and holding roughly 20% market share.
• An emergency sales push, not a science fix, saved a stalled launch. Two months after launching, only one physician was ordering the test; David describes hiring and training five additional sales reps within three weeks and having reps spend 30-45 minutes per patient explaining the test so patients would push their doctors, which lifted kit return rates to about one in five and gave them the traction needed to hire experienced sales staff.
• BillionToOne deliberately sequenced its roadmap from cheapest to most capital-intensive market. The founders say they mapped out prenatal testing, then late-stage cancer, then early-stage cancer detection as early as 2018, funding each harder, more expensive step with revenue from the last rather than raising over a billion dollars upfront for early cancer detection — the approach they credit with letting first-time, resource-constrained founders reach a "holy grail" problem others couldn't afford to attempt.
• The company staffs itself with "interdisciplinary people," not interdisciplinary teams. Ozan says small pods of one principal-investigator-level scientist plus two or three research associates own a product end-to-end and report directly to the co-founders, which he says removes bureaucratic bottlenecks and speeds iteration by an order of magnitude compared to splitting chemistry and data-analysis expertise across separate specialists.

2026-03-31 · open the original ↗

• AI agents replaced three separate legacy fraud-detection systems with one self-healing loop. Karim, Variant's co-founder, says companies used to stitch together rules engines, narrow classifiers and slow human reviewers, but agents can now read a company's own standard operating procedures and reason directly over unstructured data — and when a customer has no API, Variant's agents scrape data by operating the same legacy review dashboard a human analyst would have used.
• Reasoning across a full entity graph caught fraud isolated classifiers missed. Karim describes detecting coordinated, state-sponsored fraud rings pushing a single narrative during an election cycle, and separately catching people making credible, at-scale threats of physical violence, by having agents relate each account to every other account rather than scoring one piece of content in isolation — cases that were then referred to law enforcement.
• Coding agents let five engineers ship roughly a 25-person team's output. Karim says every engineer runs multiple AI coding agents in parallel while still reviewing all pull requests, and gives an example of a non-technical customer success manager taking a customer's feature request straight to a Cursor agent and shipping it without ever looping in engineering.
• The company nearly ended when a truck hit its only salesperson. Karim, who ran all of Variant's sales himself, describes revenue doubling month over month through mid-2024, then breaking his spine and leg in a cycling accident that hospitalized him for 10 days, leaving co-founder Michael, on an engineering-only team, questioning whether the company could continue.

2026-03-30 · open the original ↗

• The first people to live to 1,000 may already be alive today. The speaker argues AI's disruptive potential remains underpriced by the public and that brain-computer interfaces — a separate but parallel development — will reframe the human condition by around 2035, though he stops short of predicting all disease will be cured by then.

2026-03-27 · open the original ↗

• Chollet's new lab wants to replace deep learning with tiny symbolic models. François Chollet says his lab, Indium, is pursuing "symbolic descent" — searching for the smallest symbolic program that explains a dataset instead of fitting parameters via gradient descent — betting this yields models needing far less data and generalizing better, though he puts the odds of the approach succeeding at only 10-15%.
• Chollet expects task automation to arrive well before true general intelligence. He separates the industry's usual definition of AGI (automating economically valuable work) from his own (matching human data-efficiency at learning any new task), and argues today's LLM-based systems can already fully automate any domain with a verifiable reward signal — starting with code and, he predicts, mathematics next — while unverifiable domains like essay-writing will progress much more slowly.
• ARC-AGI's benchmark history shows models getting more useful without getting smarter. Chollet says reasoning models produced a step-function jump on ARC-1 in late 2024, while ARC-2 was later saturated not by higher fluid intelligence but by labs building custom harnesses that generate, verify and fine-tune on millions of synthetic reasoning chains — evidence, he argues, that recent gains reflect better training and tooling rather than a rise in raw intelligence, though he notes one YC company reached 97% on ARC-2 with this approach.
• ARC-AGI-3 tests whether an AI learns new games as efficiently as humans. Built with a hired game-development studio, it drops agents into more than 250 unlabeled environments where every exploratory action counts against an efficiency score benchmarked against human testers, and deliberately keeps its private test set conceptually different from the public one to resist the same harness-driven gaming that saturated ARC-2; Chollet predicts AGI around the early 2030s, roughly when his team releases ARC-6 or ARC-7.

2026-03-26 · open the original ↗

• Frontier models already show distinct strengths, but will commoditize within years. The speaker says Opus is a strong all-around workhorse, Codex excels at backend debugging and Gemini at frontend work, letting their product lean on each model's "spike"; they expect open source to trail by roughly three to six months and all models to converge on similar capability and pricing, leaving the advantage to whoever best serves the user.

2026-03-20 · open the original ↗

• A pancreatic-cancer enzyme became the key to a new manufacturing process. Co-founder Gora, then in medical school studying pancreatic cancer, found tumor cells locally generating roughly 50% hydrogen peroxide concentrations via a mutated enzyme; pairing that enzyme with a metal catalyst — a combination Solugen calls chemoenzymatic processing — pushed reaction yield from a typical 60% to 96% at scale.
• Solugen deliberately avoided the large-plant-first approach that had sunk earlier bio-chemical startups. Co-founder Sean says they built their first reactor for $10,000 out of PVC pipes from Home Depot and started selling small volumes of peroxide to hot tub owners in Dallas almost immediately, arguing capital constraints forced creative thinking and that deep, direct customer knowledge — not a big plant — determined what they could actually build.
• A hyper-targeted, unconventional sales tactic landed Solugen's first oil-and-gas field trial. Sean describes identifying the one person at a saltwater disposal company who controlled chemical purchasing, then buying billboards along his commute route for roughly $10,000-$15,000 so he would recognize the company by the time they called him.
• Solugen's flagship plant is its original prototype, scaled roughly 1,400x. Sean says the 60-foot bubble-column reactor at their Bioforge 1 facility is functionally identical to their original Y Combinator-era PVC reactor, just 10,000 gallons instead of 7, and that one bottle-sized dose of enzyme yields two to four tanker trucks of finished product; the company builds smaller plants near customers rather than one mega-facility to undercut larger competitors on shipping costs.

2026-03-17 · open the original ↗

• Puddle's founder claims the startup can do recursive self-improvement far cheaper than approaches that require retraining a new model from scratch. He says the standard approach costs hundreds of millions of dollars and months of effort per model; pressed that incumbents could still out-release them, he responds that Anthropic, OpenAI and Google are pursuing self-improvement too, but typically still at the expensive full-retrain-per-step level.

2026-03-16 · open the original ↗

• Emergent pivoted twice before finding its market, chasing verification, then non-technical builders. Co-founder Mukund Jha says the company started in late 2023 automating software testing, then realized solving verification lets a coding agent run autonomously for much longer, which led them to build a general coding agent that hit #1 on the SWE-bench benchmark within two months — before they abandoned an enterprise push and repositioned for non-technical users, who now make up 80% of the platform's builders across 7 million apps built in 8 months.
• Emergent built its own cloud infrastructure so build and deploy environments match. Jha says they run their own Kubernetes-based sandbox stack rather than a third-party provider specifically to avoid deployment failures, pairs a multi-agent architecture that delegates tasks like testing or API integration to sub-agents, and gives agents a long-term memory built from auto-generated "skills" distilled from past session trajectories — which he says must be generated by a separate CI/CD-style process rather than by the agent itself to actually improve performance.
• Emergent dogfoods its own product, including an agent-built Asana clone. Jha says a QA engineer's side project of prompting "clone Jira" grew into a project-management tool now used company-wide, built entirely on Emergent with no hand-edited code, saving roughly $3,000-$4,000 a month in subscription costs, and that every employee — including non-engineers — does customer support shifts to stay close to user needs.
• Non-technical domain experts are shipping real, previously-impossible businesses. Jha cites a clinical psychologist and equestrian coach in Alaska who built "EquiMind," an app combining both fields, after a Nova Scotia dev shop quoted her a price she couldn't afford; he frames custom software that used to cost around $500,000 as now buildable for roughly $5,000, and mentions one app built on Emergent has reportedly raised $4 million in funding.

2026-03-13 · open the original ↗

• YC is bringing Startup School to Bangalore for the first time, on April 18. The one-day event gathers top Indian builders and hackers and features founders of YC-backed Indian companies including Razorpay, Meesho, Groww and Emergent.

2026-03-09 · open the original ↗

• Science's retinal implant produced the first-ever coherent "form vision" from a prosthesis. Max Hodak, the company's founder, says their clinical trial — published in the New England Journal of Medicine, covering more than 40 patients across 17 European sites — let people who'd been unable to see faces for a decade read letters on an eye chart, which he attributes to stimulating the retina's bipolar cells rather than the optic nerve's ganglion cells, preserving compression the retina had already done; an earlier device from a company called Second Sight, which stimulated ganglion cells directly, produced only meaningless flashes that patients' brains never assembled into images.
• Adult brains stay far more plastic under feedback than commonly believed. Hodak says early motor-decoding experiments deliberately used fixed, arbitrary mappings between a few neurons and cursor movement, and patients' brains learned to control the mapping regardless — evidence, he argues, that plasticity persists throughout adulthood — but contrasts this with adults born blind from congenital cataracts whose sight was surgically restored decades later and whose brains, past a critical developmental window, could never learn to interpret the input, with some patients dying by suicide.
• Science tested every stimulation approach before choosing electrical bipolar-cell stimulation. Hodak describes exploring a 2-by-2 matrix of choices — stimulating bipolar cells versus optic nerve ganglion cells, done electrically versus via optogenetics — and rejecting ganglion-cell stimulation because it required an impractical million-degree-of-freedom calibration per patient; the company still develops an in-house optogenetic alternative sensitive enough to activate under ordinary indoor lighting, which Hodak estimates is 5 to 7 years from clinical use.
• Science's "biohybrid" interface grafts living neurons onto the brain itself. Hodak says the approach — inspired by the brain's own corpus callosum and by conjoined twins in Canada who share elements of conscious experience through one connecting nerve bundle — seeds a device with hypoimmunogenic, stem-cell-derived neurons that biologically wire into the patient's brain without inserting wires or genetically modifying the patient's own cells, avoiding the largely irreversible risk he says comes with gene-therapy-based approaches like optogenetics.

2026-03-06 · open the original ↗

• LLM-trained design clichés are homogenizing startup websites overnight. Raphael Schaad, YC visiting partner, argues purple gradients, cursor-trailing lines, and scroll-triggered fade-ins spread instantly once one well-linked site enters LLM training data — a process that used to take months of manual copying before AI design tools existed.
• Hover effects on AI-built sites frequently hide rather than reveal. Menus and buttons on sites like Nunu (new.ai) fade or vanish on hover instead of "popping," which Schaad calls an anti-pattern since browsers already signal clickability and hover has no mobile equivalent, making it a poor place to hide essential functionality.
• Scroll-jacking and cursor-chasing buttons distract more than they convert. Garry Tan and Schaad found Get Crocs's roaming call-to-action button and Zarna's hijacked, molasses-slow scroll pulled their attention away from the actual value proposition rather than toward it.
• Founders must edit AI output, not accept it as final. Schaad points to undetected bugs shipped on one-shotted pages — a broken selection state on Build Zero, semi-clickable elements on Zarna — and advises starting from a company's own brand and color palette rather than whatever the AI defaults to.

2026-03-05 · open the original ↗

• YC's Startup School returns for two days in San Francisco, July 25–26, with speakers including Jensen Huang, Sam Altman, Alexander Wang and Jeff Dean. The event targets top undergrads, grad students and researchers in CS, ML, applied math and engineering, offering direct access to YC partners and alumni; YC says over 100 attendees from last year's event are now YC founders.

2026-02-27 · open the original ↗

• Poetic builds AI harnesses, not fine-tuned models, atop frontier LLMs. Ian Fisher, co-founder and co-CEO, argues fine-tuning becomes worthless once a new frontier model ships, since firms have sunk hundreds of millions into a now-obsolete base; his "meta system" instead auto-generates reasoning harnesses that stay compatible and keep improving for free as underlying models advance.
• A harness beat Google's top model on ARC-AGI for less money. Fisher says that two days after Gemini 3 Deep Think posted 45% on ARC-AGI v2 at roughly $70 per problem, Poetic's system — built on the cheaper Gemini 3 Pro — reached 54% at $32 per problem.
• Poetic also edged Anthropic's Opus 4.6 on Humanity's Last Exam. Fisher reports Poetic scored 55% on the 2,500-question expert benchmark, versus Opus 4.6's 53.1% announced the previous week, at an optimization cost he says was under $100,000, done by a seven-person team.
• Automated reasoning strategies, not better prompts, drive the biggest gains. Fisher cites earlier DeepMind work where manually engineered reasoning strategies lifted Gemini 1.5 Flash's score on a hard task from 5% to 95%; Poetic's meta-system now discovers such strategies itself rather than relying on hand-tuned prompts or curated datasets.

2026-02-24 · open the original ↗

• The "software engineer" title will fade as coding becomes a universal skill rather than a separate job. The speaker says every function on their team already codes — PMs, designers, the engineering manager, even finance — and expects engineers' work to broaden into writing specs and talking to users as this pattern spreads industry-wide.

2026-02-21 · open the original ↗

• Agents now pick dev tools by parsing documentation, not word of mouth. YC partners argue this is reshaping dev-tool go-to-market, citing a surge in Postgres databases created over the past year as agents building apps default to Supabase, whose documentation they say is simply the best-written for an LLM to parse.
• Resend rebuilt its docs to be agent-readable, and it worked. The email-sending company's founder noticed over a year ago that ChatGPT had become a top-three inbound channel, so he restructured docs into bullet-point Q&A with embedded code snippets; partners contrast this with SendGrid's support-ticket-gated docs, which they say leave it invisible to agents.
• Documentation infrastructure is becoming a must-have, not a nice-to-have. Partners point to Netlify, built on better API documentation, as newly positioned to capture "exponentially more agents making exponentially more decisions" than developers ever did — meaning even a small documentation improvement could meaningfully move a dev-tool company's business.
• Agent Mail's agent-only inboxes took off once OpenClaw exploded. The YC company built email accounts designed for AI agents because Gmail and other providers deliberately obstruct automated sign-ups to fight spam; partners speculate this points toward a fuller "agent economy" — even agent-native phone numbers — running alongside, and possibly eventually distinct from, human commerce.

2026-02-20 · open the original ↗

• Claude Code's creator didn't know at first that he had something, then became consumed building it. He describes working every night and weekend for three straight months without a vacation once he sensed "this is going to be a thing," even though it couldn't yet code; he says Anthropic's internal principle — build for the model six months out, not for today's — remains his advice to founders building on LLMs.

2026-02-17 · open the original ↗

• Claude Code began as an accidental internal prototype, not a plan. Boris Cherny, its creator, says he built a terminal chat wrapper just to learn Anthropic's API, added tool use experimentally, and only realized it mattered when a coworker used it to code the next day and CEO Dario Amodei asked why internal usage looked mandated — it wasn't.
• Anthropic builds for the model six months out, not today's. Cherny says Claude Code is entirely rewritten every few months because scaffolding built to patch a model's weaknesses becomes wasted work once the next model closes the gap for free; his advice to founders is to build for the frontier the model hasn't reached yet.
• Claude Code reportedly reshaped Anthropic's own engineering productivity. Cherny cites an internal measure — pull requests, cross-checked against commits — showing productivity per engineer up roughly 150% since the tool's release, with his own coding now 100% Claude Code and zero hand-written lines since Opus 4.5.
• He predicts the title "software engineer" fades as coding gets solved. Cherny forecasts that as coding becomes "generally solved," roles will shift toward specs and user conversations across every function — he says Anthropic's PMs, designers, and even finance staff now code — while a riskier further frontier is ASL4-level recursive self-improvement, which he says Anthropic is actively working to make safe.

2026-02-14 · open the original ↗

• The leanest startups now automate every internal function, not just one. Garry Tan calls these "20X companies," extending Parker Conrad's "compound startup" idea from product breadth to internal operations — covering code, support, marketing, sales, hiring, and QA — which keeps payroll and headcount down as revenue scales.
• One customer-service startup beat 20-times-larger rivals for a major account. GigaML's founders say they closed DoorDash as a customer with just four to five engineers against competitors with roughly 100x the headcount, crediting their internal agent Atlas — which can use browsers, edit policies, and write code — with doubling or tripling each engineer's scope.
• One interface let a psychiatry network 4x patients with flat headcount. Legion Health built a single internal tool giving its care-ops team instant access to patient history, scheduling, and insurance codes; it says this let revenue and patient volume grow 4x in a year while staffing stayed at one clinical lead, one patient-support person, and one billing person.
• A 12-person firm skips hiring by turning workflows into agents. Feathr, competing against accounts-receivable incumbents dating to 2006, has employees document their manual tasks so it can build custom agents for each; it says this let it avoid hiring a dedicated designer altogether, using an AI design tool instead for all front-end work.

2026-02-12 · open the original ↗

• OpenClaw's edge is running locally on a user's own computer rather than in the cloud. Speakers say that access lets it control anything connected to the machine — an oven, a Tesla, lights, a Sonos, even a smart bed — and describe it searching a user's files unprompted to compile a surprising narrative of their past year from forgotten old audio recordings.

2026-02-11 · open the original ↗

• AI-native hedge funds will emerge by inventing new strategies, the way quant funds did in the 1980s, not by bolting AI onto old ones. A former quant researcher recounts being ignored after asking compliance to approve ChatGPT use at a major fund, and argues the opportunity is swarms of AI agents combing 10-Ks, earnings calls and SEC filings to generate and trade on genuinely new strategies.

2026-02-10 · open the original ↗

• Stablecoins sit in a regulatory middle ground that opens room for new DeFi-meets-TradFi financial products. The speaker cites the GENIUS and CLARITY Acts as positioning stablecoins as compliant yet crypto-native, enabling services like yield-bearing accounts or tokenized real-world-asset access without forcing a choice between limited traditional products and unregulated crypto risk.

2026-02-09 · open the original ↗

• Spatial reasoning, not just language, is the next frontier foundation-model category. The speaker argues current systems only handle basic depth and relationship tasks and can't robustly reason about 3D manipulation or mental rotation, and that a company treating geometry as a first-class primitive rather than a language add-on could become a foundation-model player on the scale of OpenAI or Anthropic.

2026-02-09 · open the original ↗

• Government needs AI tools to handle a coming surge of forms still processed by hand. The speaker points to Estonia as an early model of digital government and argues that, despite how hard it is to sell to government, landing a first customer tends to produce sticky contracts that expand over time.

2026-02-08 · open the original ↗

• Slow American mill lead times are ceding the metal market to cheaper, faster overseas competitors. The speaker says an aluminum plate can take up to 8 weeks and stainless steel for aerospace and defense up to 8 months from a US mill, and argues new domestic mill capacity is needed to support manufacturers like RMFG, Forge Automation and Atomic Industries.

2026-02-07 · open the original ↗

• OpenClaw's edge is running locally with full control of your machine. Peter Steinberger, its creator, says cloud-only assistants are inherently limited while a local agent "can do every effing thing" — controlling smart devices, files, and any application the way a human user could.
• A model improvised a voice-transcription pipeline it was never built for. Steinberger recounts sending OpenClaw a voice message while traveling with poor internet in Marrakech; with no transcription feature built, the model identified the audio format, converted it with FFmpeg, found a stored OpenAI key, and called the Whisper API via curl — replying in about nine seconds, which he says convinced him coding models' problem-solving generalizes to real-world tasks.
• He predicts 80% of apps disappear as agents manage data directly. Steinberger argues most apps just manage data an agent could handle more naturally — tracking meals from a photo, adjusting a gym schedule, setting reminders — leaving mainly apps built around physical sensors, while large model companies keep a moat through memory and context locked inside their own data silos.
• He skipped git worktrees and MCP support to cut complexity. Steinberger runs many parallel plain checkouts of the same repository instead of branches, and built no native MCP integration into OpenClaw, instead exposing MCPs as command-line tools through a separate converter — which he says needs no restart and scales better than the MCP approach used by Codex or Claude Code.

2026-02-06 · open the original ↗

• LLM training tooling hasn't kept pace with the field and still breaks constantly. Drawing on three years training diffusion and language models, the speaker describes losing time to broken SDKs, GPU instances that fail only after a half-hour setup, and buggy open-source tools, and calls for better training APIs, large-dataset management tools and ML-native dev environments as post-training and model specialization grow in importance.

2026-02-06 · open the original ↗

• Government fraud detection tooling is decades out of date. YC says most systems sold to agencies are rules engines from the mid-2000s with bolted-on machine learning, generating thousands of context-free alerts that humans end up ignoring; Medicare alone is cited as losing tens of billions of dollars a year to improper payments.
• Qui tam whistleblower suits are the one mechanism that already claws back government money at scale, but the process is entirely manual. An insider tips off a law firm, which then spends years pulling documents and building a case; YC argues this should be software that parses messy PDFs and medical records, traces opaque corporate structures, and outputs lawyer-ready case files.
• YC advises selling to intermediaries, not agencies, and wants founders who've done the work. Recommended first customers are whistleblower law firms, inspector generals, and state AG fraud units rather than big government agencies directly, with a founder profile that includes former False Claims Act counsel, healthcare compliance leads, or procurement auditors.

2026-02-06 · open the original ↗

• Claude Code's edge comes from splitting work across sub-agent context windows, reflecting a real architectural fork between labs. Kelvin French-Owen, who built Codex at OpenAI before founding Segment, says Claude Code spawns "explore" sub-agents running smaller models to search the file system in their own context windows before reporting back — while Codex instead runs periodic compaction to sustain much longer autonomous jobs, which he traces to Anthropic building "tools for humans" versus OpenAI pushing toward fully autonomous, longer-horizon models.
• Context window size is still the top bottleneck, even for expert users, so managing it is an active skill. French-Owen says he manually clears Claude Code's context once it crosses roughly 50% token usage to avoid what he calls the "dumb zone" — a concept he credits to Dex at the YC company HumanLayer — where output quality degrades as the window fills; he also describes founders planting an odd "canary" fact early in a session to test when the model starts forgetting.
• Building toward full test coverage sharply increased his coding speed. French-Owen says he worked with little to no testing for his first few days using Claude Code, then dedicated a day to reaching 100% test coverage, after which he rarely needed to manually verify changes because the agent's own test runs caught regressions — paralleling what the hosts say other teams have found doing test-driven prompt engineering outside of coding.
• Bottoms-up, engineer-first distribution is reshaping which tools and even architecture choices win. French-Owen and the hosts argue coding CLIs succeed because individual engineers can download and use them without enterprise approval, and that open-source projects like Supabase benefit disproportionately because both LLM recommendations and coding agents reading source code directly favor visible, well-documented code over proprietary alternatives.

2026-02-04 · open the original ↗

• Real-time AI guidance, not automation, is the near-term opportunity in physical work. For fields like field services, manufacturing, and healthcare where AI can't yet act in the world, the pitch is a worker wearing a camera while an AI sees what they see and talks them through the job step by step — "turn off that valve," "use the half-inch wrench," "that part looks worn, replace it."
• The model is instant competence rather than months of training. With AI coaching in real time, YC argues a worker can be effective immediately and pull in new skills on demand instead of needing years of apprenticeship.
• Three converging factors make this timely now: multimodal models that can reliably see and reason about real-world situations, hardware already in place (phones, AirPods, smart glasses), and skilled-labor shortages that make the problem economically urgent. YC suggests three possible approaches — selling to companies with existing workforces, building a full-stack vertical (e.g. HVAC repair or nursing), or a platform letting anyone become a skilled worker.

2026-02-03 · open the original ↗

• AI lets agencies flip from selling services to selling finished products at software margins. The speaker argues agencies can now use AI themselves to produce finished work upfront — a design firm pitching finished creative before a contract is signed, an ad agency generating video without a physical shoot, a law firm drafting documents in minutes — letting them charge far more and scale past the labor-bound limits of today's fragmented agency markets.

2026-02-02 · open the original ↗

• AI coding tools solve building software, not deciding what to build. YC argues Cursor and Claude Code are good once it's clear what needs building, but the harder, more important work — talking to users, understanding markets, synthesizing feedback — still lacks an AI-native system covering the full product-discovery loop.
• The pitch is a tool that turns customer data directly into a proposed feature. Upload customer interviews and usage data, ask what to build next, and get back a feature outline with a rationale grounded in that feedback, plus proposed changes to the UI, data model, and workflows.
• The tool would also hand off directly to coding agents. It would break the proposed changes into development tasks sized for a coding agent to implement, reflecting YC's view that as agents take the first pass at implementation, how teams define and communicate what to build needs to change.

2026-01-30 · open the original ↗

• The old YC homepage read as generic and underplayed its own founders. YC's designer says the previous site felt like "a B2B SaaS template" with vague, unquantified claims ("we improve the success rate of our startups" without a number) and showed portfolio company logos without crediting the founders who built them — a gap the redesign set out to fix.
• The redesign leans on storytelling rather than a sales pitch, drawing directly on Paul Graham's essays. The team studied PG's writing and pulled the word "formidable," which he and Jessica Livingston used for extraordinary early founders, into the new hero section, complete with a PG-style footnote defining it.
• Founders, not YC, are put front and center. The new site shows before/after images of funded founders, side-by-side quotes stitched together from interviews with recent-batch founders about their YC experience, and photos of YC partners from when they themselves went through a batch — reinforcing that partners are ex-founders. The team deliberately kept most of the original circa-15-year-old PG homepage copy verbatim, updating only details like funding cadence and the San Francisco location.
• The site was built by prompting Opus 4.5 in Cursor rather than designing static frames in Figma. The designer says this let the team prototype interactive sections live and iterate toward outcomes — like a scrolling before/after carousel — they don't think they'd have arrived at in a static tool, freeing their time to focus on storytelling and interaction design instead of manual implementation.

2026-01-29 · open the original ↗

• Generic illustrations of a product are a recurring failure across the sites reviewed, and screenshots or a short video would fix most of them. Framer co-founder Jorn van Dijk repeatedly tells founders (Lumari, the presentation tool "a lie") that vague icon animations don't tell a first-time visitor what the product does or looks like, and that a real screenshot or two-minute walkthrough would do more to earn trust than an abstract hero animation.
• Asking for a demo or signup before showing any value creates unnecessary friction. Van Dijk argues a "book a demo" button is a big ask before a visitor understands what they're buying, and for the presentation tool competing directly with free Google Slides, he says letting people try the product immediately — even a prompt box on the homepage — matters more than gating access behind a signup wall.
• Visual style needs to match what a site claims to be, or it undercuts the message. Van Dijk flags that Juicebox's black-and-white, "win the talent war" aesthetic feels disconnected from being a people-focused hiring product, and that Leaping AI's single-phone, consumer-style hero visual works against its stated positioning as an enterprise call-center replacement.
• Small inconsistencies compound into distrust of a site's bigger claims. Reviewing The Hog, Van Dijk points to uneven letter spacing, a fully capitalized customer quote with no named company attached, and a shared toggle widget that behaves two different ways on the same page, arguing that a visitor who notices these details starts to doubt the numbers and testimonials elsewhere on the page too.

2026-01-23 · open the original ↗

• They set out to build a fast-turnaround, fully reusable orbital rocket. The founders say the industry norm is to build and launch a rocket once, lose it in the ocean, and never see it again; they wanted a vehicle that reaches orbit, flies back, and repeats the cycle "again and again," a problem they felt nobody else was attacking with rigor.
• Launching the company felt like an irresponsible, highest-stakes bet. One of the two co-founders calls starting from zero, with no one else in the field beside them, the hardest personal decision they've made, reached only once they concluded the idea "had to be tried" regardless of whether it succeeded.

2026-01-22 · open the original ↗

• Diffusion is a general-purpose way to learn any data distribution by training a model to reverse noise, and it now spans far more than images. Visiting partner Francois Chaubard cites DeepMind's Nobel Prize-winning protein-folding work, the "diffusion policy" approach used for robotics and self-driving, and Google's GenCast, which he calls the most accurate weather forecasting system in the world, as the same core mechanism applied to different domains.
• Progress came from making the training objective easier to learn, not from a new core idea. Chaubard traces a line of papers that moved from predicting the noised data itself, to predicting the added error, to predicting a "velocity," each step making the math simpler and driving the Fréchet Inception Distance quality metric down from about 222 in the original 2015 paper to roughly 8-10 in modern models.
• Flow matching, from Meta researcher Yaron Lipman, collapsed diffusion training into a straight line between noise and data. Chaubard demonstrates it as roughly five lines of training-loop code that work unmodified whether the underlying model is generating images, proteins, DNA, or robot trajectories, calling it one of the more profound simplifications in recent machine learning.
• Chaubard's advice splits by role: model-trainers should treat diffusion as a default tool, and everyone else should update on how fast it's improving. He argues anyone training their own models should seriously evaluate diffusion regardless of application, and that non-trainers should expect roughly the same ~1,000x quality jump seen in image generation over five years (Midjourney to Sora/Veo/Flux) to extend into proteins, robotics policies, and self-driving as the technique keeps getting simpler.

2026-01-15 · open the original ↗

• The AI economy stabilized into distinct model, app, and infra layers. A YC partner says 2025's biggest surprise was how much calmer things got compared to late 2024's rapid shifts, with a clear playbook now for building AI-native companies atop the models; he reads the stability as downstream of incremental model improvements rather than any single disruptive breakthrough this year.

2026-01-14 · open the original ↗

• Finding first users is a search problem, not a persuasion problem. Almost no one wants to be a startup's first customer, so the goal is finding the rare people who love trying new things or have a burning need — illustrated by the speaker paying and becoming the first customer of an unknown startup within three days because it solved a specific billing/API problem.
• Charge real money early and use personal, targeted outreach. Early adopters and people with a burning problem aren't price-sensitive, and paying customers give sharper feedback than free ones; finding them looks like cold email or direct outreach, not mass channels like a billboard. The advice also includes launching early, studying early users like an anthropologist, and not fearing churn since a lost early user is low-stakes and easily replaced.
• Personal software budgets are small, which shapes where AI founders sell. The speaker cites their own ~$150/month personal software spend versus multiple corporate tools each costing more than that alone, arguing this gap is why many AI founders target prosumers, businesses, or high-advertising-value users like doctors rather than pure consumers.
• Build a "minimum evolvable product," not just an MVP, because early adopters steer what the product becomes. Using Tesla's Roadster as an example, the speaker argues early buyers who wanted an impractical, fast, tech-forward car — not comfort — are why the mass-market Model Y still has under-3-second 0-60 acceleration and worse suspension than a Toyota; product evolution is path-dependent on who adopts it first.

2026-01-08 · open the original ↗

• Stoke Space is targeting full rocket reusability, including the second stage that every other reusable rocket still throws away. A co-founder explains that industry has already made first-stage reuse routine, scaling launches from roughly 10-20 to about 150 a year, but the upper stage still burns up structurally around 2,700°F re-entering at 17,000 mph; Stoke's Andromeda capsule survives this with a heat shield cooled by liquid hydrogen and 24 small thrusters for landing.
• The founders quit stable jobs at Blue Origin in 2019 to found the company on real personal risk. One co-founder describes it as "wildly irresponsible" and his hardest personal decision, having a 3-month-old at home; they set a 6-month deadline to find traction, got their first funding check right around that point, just as COVID hit, and built their first prototype engine in a shipping container in a backyard.
• YC's network was central to raising capital in a market skewed toward SaaS. The founders say they wildly underestimated how hard fundraising would be as first-time hardware founders with no investor connections, and credit YC with teaching them the fundraising process and adding momentum to the round; the company has raised about $990 million to date.
• Fast iteration is treated as the core competitive advantage, enabled by in-house manufacturing and custom software. By building nearly every part themselves rather than depending on outside suppliers, a co-founder says they cut a test-fail-rebuild cycle from about a month down to a day or two; the company also built its own operations software, called Bolt Line, to track part history and maintenance as it scales toward its first orbital launch, planned for later this year from a new site at Cape Canaveral.

2025-12-22 · open the original ↗

• Anthropic overtook OpenAI as the top model choice among YC's Winter 2026 batch, a reversal from OpenAI's earlier dominance. Partner Diana says OpenAI held 90%+ share in early batches and Anthropic hovered near 20-25% through 2024 and early 2025 before surging past 50% in the last few months; she attributes it partly to Anthropic making coding performance an internal north-star eval, feeding the batch's wave of coding-agent and vibe-coding startups. Gemini has also climbed from roughly 2-3% to about 23% over the same period.
• Founders increasingly treat models as commodity, interchangeable inputs rather than picking one loyally. Partners describe series-B-level founders building an orchestration layer that swaps models per task based on proprietary evals — one startup reportedly uses Gemini 3 for context engineering and feeds the result to OpenAI for execution — comparing it to how chipmakers Intel and AMD get swapped out as each releases a better architecture.
• The "AI bubble" framing is reinterpreted through economist Carlota Perez's installation-versus-deployment model of tech revolutions. Partners argue the current heavy infrastructure capex — gigawatt data centers, and even fringe bets like space-based data centers and fusion-power startups such as Zephyr Fusion — sits on the balance sheets of big companies, not startups, so a slowdown there wouldn't strand founders; the resulting compute glut is framed as setting up the "deployment phase" for application-layer startups, comparable to how excess telecom fiber from the dot-com buildout later enabled YouTube.
• Early fine-tuning-based moats are proving fragile against leapfrogging frontier models. Partners note a YC company that beat GPT-3.5 through fine-tuning was later "blown out of the water" once GPT-4.5 and 5.1 shipped, while newer domain-specific models fine-tuned with reinforcement learning on open-source bases can still beat general models on narrow tasks — one YC healthcare startup is cited as outperforming OpenAI's models on healthcare benchmarks with only an 8-billion-parameter model.

2025-12-17 · open the original ↗

• ARC Prize defines intelligence as the efficiency of learning new skills, not the ability to execute known ones. Its president explains the foundation built on François Chollet's 2019 definition, contrasting it with benchmarks like MMLU that chase harder "PhD-level" problems; ARC tasks are deliberately kept solvable by ordinary people, testing whether a system can generalize rather than just score high.
• Reasoning models produced a step-change on the ARC-AGI benchmark that pure pretraining never did. GPT-4's base model scored only about 4-5% as of 2024, but o1 and o1-preview jumped to roughly 21% once reasoning was introduced, which the foundation says is why OpenAI, xAI, Google, and Anthropic have all reported ARC-AGI scores for their frontier releases (Grok 4, Gemini 3 Pro/Deep Think, Opus 4.5) over the past 12 months.
• The president warns against mistaking narrow RL-environment scoring for real progress. He argues that building a custom RL environment to ace a specific benchmark doesn't generalize the way humans do without one, calling it "whack-a-mole," and says ARC guards against this with a hidden test set and an emphasis on novel problems.
• ARC-AGI-3, due next year, will be an interactive benchmark with no instructions and will score efficiency, not just accuracy. Test-takers must infer the goal of roughly 150 game-like environments purely from action and feedback, and performance will be normalized against average human action-counts rather than raw success rate. The president maintains that even a hypothetical 100% score would be necessary but not sufficient evidence of AGI, prompting further scrutiny of the system's remaining failure points rather than a declaration.

2025-12-10 · open the original ↗

• PostHog's breakout product came directly from pain its founders caused themselves through repeated pivots. CEO James Hawkins says after a failed sales-territory tool (14 of 15 interested prospects never even clicked the signup link) and a failed developer-survey product, the team kept having to re-set-up product analytics after every pivot — leading them to build the self-hosted, open-source analytics tool that finally worked, on the insight that technical users make better early customers because "what they say is close to what they'll do."
• Discipline, not inspiration, drove the search: charge money early and treat non-payment as your clearest signal. Hawkins says the team committed fully to roughly one idea every five to six weeks, meeting every interested prospect in person, and found that struggling to sell a $50/month product was as hard as the enterprise deals he used to close — proof the idea itself wasn't solving a real problem, since he knew they weren't holding back effort.
• Fundraising flipped from brutal to easy over five years. The 2020 seed round coincided with COVID gutting VC interest overnight, forcing PostHog to raise via $5,000 angel checks after meeting roughly 160 firms; the recent $75 million round at a $1.4 billion valuation — making PostHog YC's newest unicorn — came together after talking to essentially one investor who'd offered to lead it preemptively.
• The current bet is building AI that automates roles which already have product-market fit inside companies, starting with product management. Hawkins describes a "pull-based" tool in development that scans session recordings, analytics, error tracking, and LLM traces across PostHog's 15-16 products to open pull requests fixing issues automatically while engineers sleep, reasoning that if a human product manager role sells, an AI version of it should too.

2025-12-05 · open the original ↗

• Many good startup ideas live in legal gray areas. He cites OpenAI's crawling of the entire web without permission as arguably fair use or "massive copyright theft," and recalls Lyft's founders fearing jail the week before launch, since what became Lyft and Uber was then effectively illegal.
• Laws often get rewritten once consumers clearly benefit. He argues that finding laws written before a big tech shift, laws that no longer reflect reality, can be valuable for founders, since the legal environment tends to bend once it's clear the end user wins.

2025-11-13 · open the original ↗

• Orbital data centers could remove Earth's energy and water constraints on AI compute. Philip Johnston, co-founder of StarCloud, argues data centers in space draw constant solar power, radiate heat via infrared into deep space instead of evaporating fresh water, and can scale without land or grid limits, ultimately competing on cost with the largest terrestrial data centers.

• A large, low-cost deployable radiator is StarCloud's core technical bet and IP. Johnston says the main public criticism, from a viral post by Andrew McCallip, is that dissipating heat needs impractically large surface area; he counters that co-founder Ezra, a satellite-structures PhD, is leading half the engineering team to just build that large radiator.

• StarCloud pivoted from beaming solar power to Earth to sending data centers into orbit instead. Johnston says the team's original space-based-solar concept loses 95% of energy in transmission and only breaks even near $50/kg launch cost; recalculating for orbital data centers instead put breakeven near $500/kg, much closer to today's falling launch prices.

• StarCloud built and launched its first satellite in 15 months, versus roughly 4 years for prior space startups. The company, backed by YC after a third application attempt (under its earlier name Lumen Orbit), launched its StarCloud 1 satellite with an Nvidia H100 GPU on November 2nd and plans a second, Blackwell-equipped launch in October 2026.