We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Chip startups keep drawing competitive rounds
Benchmark won a contested bid to lead a new round in Tendrils Compute, an early-stage chip startup based in Cambridge, UK, and led by Nils Cremer. People familiar with the deal say Tendrils is already discussing a quick follow-on that could value it above $1B . The technical bet is unusual: general-purpose chips built on interaction nets, a graph-based computing model. The pitch is that CPU speed becomes a bottleneck as agents chain inference steps and tool calls . Benchmark has done this before. It co-led Cerebras's 2016 Series A and earned about 20.5x when Cerebras went public .
Volantis announced an $88M Series A for optical interconnects that target AI's memory bottleneck. Its claims include inference at up to 10,000 tokens/sec per user on models larger than 10T parameters. It says its next chip iteration has already taped out . Newcomer lists other recent chip rounds:
- Etched: $700M at a $21B valuation
- Fractile: in talks at $6.5B after an Anthropic supply deal
- OLIX: $312M at $3.3B
- Euclyd and Delos Data: nine-figure rounds
There is a counterweight. Cerebras traded as high as $386 after listing but now sits close to its $185 IPO price .
Cheap "decision models" are becoming their own category
Perplexity is open-sourcing pplx-decider-27b, a multimodal model that outputs a probability distribution over a fixed set of answers instead of free text. It is served through a Decisions API at $0.04 per million input tokens with free output, and Perplexity says the price will drop further . Cloudflare open-sourced its first homegrown decision models, clef and clef-flash, on Workers AI the same day . Harrison Chase described the use case: cheap typed answers for small calls inside a harness, such as routing, approvals and judging, with a big model handling everything else .
Jev from TypeSafe AI started this category. On 20VC x SaaStr, the hosts said TypeSafe raised a $40M seed at $200M and is now in talks to raise more than $1B at a valuation above $10B. Jason Lemkin cited Jev at 17% of OpenRouter traffic . With two open-weight entrants priced about the same, the main question for anyone underwriting TypeSafe at that valuation is how defensible it is.
LangChain published data supporting routing. In a 973-thread A/B test, its Open SWE harness either routed each task to the cheapest adequate model tier or always used GPT-6 Astra. Merged-PR rates were 29.2% routed and 27.3% control, a difference that was not statistically significant (p=0.49). Median cost per thread fell 64%, from $2.61 to $0.94 . A fast-model-only arm was stopped within a day because output quality was too low .
The 20VC x SaaStr panel on venture now
Benchmark's Jack Altman joined the panel because Benchmark co-led Instinct's $1B round at $10B. That round came 33 days after Instinct was raising at $2.5B. The founder says Instinct is approaching $1B in annual transactions, more than half of them travel . Main threads:
- Neolab exits. AMD's $8.2B all-stock purchase of World Labs, about 2.5 years after founding, was described as the first big neolab exit. One partner's tally counts 102 neolabs that have raised more than $70B. Altman's answer was that about 10 companies can now do $10B acquisitions, which he called faster than an IPO .
- Seed sizing. Altman called $3–6M checks for 8–15% ownership "broken in many lanes." His reasoning: labs need $200M or more, and well-networked founders skip straight to $50M rounds. Lemkin's view is that $2–3M is still the natural seed size when investors aren't competing to fund you .
- Inference and open weights. Modal tripled its valuation to $15B and Baseten is in talks at $26B. Altman called inference an index bet on everything outside the labs . Lemkin thinks open-weight share has peaked. He cites closed-model price cuts and says he couldn't find an enterprise at Dreamforce willing to run mostly China-origin open weights. Altman expects more enterprise post-training on open weights, but thinks compute ownership decides the outcome .
- IPOs. Oura pulled a $15.62B IPO two days before pricing despite reports the book was oversubscribed. One panelist pointed to Forerunner's plan to sell its entire stake, which makes sellers very price-sensitive. Altman noted no AI-native application company has gone public yet, and expects several in 2027 if markets hold .
SaaStr's reading of a16z's market deck adds data on how spread out outcomes have become:
- Under-one-year-old B2B companies are growing more than 500% year over year. Mature B2B companies are at about 24% .
- 55% of US unicorns have less than two years of runway .
- For the 2024 vintage, top-decile VC net IRR is 40.5% against a -3.3% median. Median DPI is 0.00x for every vintage since 2021 .
Early-stage rounds and teams
- doxxnet (Barrett Lyon, co-founder of Prolexic) raised $38M led by a16z. The product is private networks for people and their AI agents with no server in the middle, and the company says it has blocked 38M+ threats since December . a16z's thesis is that agents widen the personal attack surface .
- Halluminate builds reinforcement-learning environments for knowledge work beyond coding, starting with finance. With fewer than 10 people, it works with four of the top five closed US labs and recently raised a $30M Series A. Its founders argue verification matters more than task volume .
- Harmonic's Hot 25 ranks Resolve AI first for the third time. The Latent Co is the top newcomer at #2, with Prime Intellect, Strala AI and Trajectory Labs also new to the list .
- YC promoted founders Vivian Shen (Juni, Acely) and Raphael Schaad (Cron) to General Partner . Garry Tan says the partnership now has 19 members, the largest in its history .
- Imbue Studio went to waitlist. It is pitched as a personal computer you shape by describing what you want, open source under FSL, runnable locally, and able to switch between AI providers .
- AMI Labs. Yann LeCun says it is still in research with no short-term product plan, has 50–60 staff across Paris, New York, Montréal and Singapore, plans more funding rounds to pay for compute, and does not intend to be acquired .
Models and research
Google DeepMind's Gemini 4 Argon is available first only to government users and trusted cyber defenders . Artificial Analysis scores it level with GPT-6 Astra, at $1.99 per task versus $3.26. The savings come from introductory pricing, not efficiency: Argon uses 62K output tokens per task against Astra's 27K . Vals lists 262K maximum output tokens against Google's 1M claim .
LlamaIndex says its Extract v2.5 document-extraction agents beat Opus 5.5 and GPT-6 Sol while costing 30% to 4x less . arXiv now limits each submitter to two submissions per calendar month .
- Instinct closed a $1B Series C at a $10B valuation. The investment thesis is consumer agents that act across the third-party internet, where independent companies and model labs could both win; the key test remains whether people use them regularly and trust them to act autonomously.
- AMD’s reported $8.2B all-stock acquisition of World Labs came 2.5 years into the company’s journey. Panelists cited about 10 potential buyers capable of making $10B acquisitions and expected more large deals over the next 6–12 months if the market holds. They also cited 102 neo-labs with $70B+ raised and argued that proven technical credentials matter for becoming an attractive acquisition target.
- The discussion cites Modal tripling to $15B and Base10 talks at $26B, and describes inference as a strong way to invest beyond the model labs. The case rests partly on unsustainable model costs and tasks where additional intelligence may add little value. Panelists disagree on open weights: one expects their share to fall, citing labs’ ability to cut prices and enterprise reluctance toward China-origin models; others point to more U.S. enterprise fine-tuning and self-hosting, with compute ownership also potentially decisive.
- Panelists describe AI venture as unusually high-variance, with elevated valuations and traction, skewed returns, and difficult position sizing; they emphasize retaining enough shots on goal. Funding needs are bifurcating: estimates cited are a few hundred million dollars to enter model-lab development and $500M for semiconductors, while some software companies can get far with $2M–$3M; traditional $3M–$6M seed rounds for 8%–15% ownership are described as broken or absent in many lanes.
- Aura withdrew its planned $16B IPO shortly before pricing; panelists did not establish the reason, but discussed its large secondary component as a possible source of seller price sensitivity.
- VC fund strategies are diverging: Bessemer announced $5.75B in fresh capital, including a $1.75B seed fund, while NFX said it would invest GP capital rather than raise new LP capital for new funds.
- AI is affecting Kickstarter on both the founder and product sides: CEO Everett Taylor said tech layoffs—including some he linked to AI—are pushing experienced workers toward entrepreneurship, AI is making it easier to start companies, and AI-powered hardware is an “exploding” category with projects raising millions.
- Taylor presents reward crowdfunding as a non-dilutive source of capital in exchange for rewards; he says Kickstarter’s Design & Technology category grew 70% last year and its audience of technology early adopters can provide campaigns with customers, feedback, advocates, and sales proof that may improve leverage in later VC fundraising.
- Kickstarter is not an idea-only funding route: Taylor says hardware founders need upfront investment for a working prototype. He advises founders to establish some product-market-fit signal, prepare marketing and video, and have manufacturing and fulfillment ready; B2B campaigns can work when they offer engaging rewards, but he says they are generally a weaker fit for multi-million-dollar raises than B2C products.
- Taylor contrasts reward crowdfunding with equity crowdfunding: Kickstarter backers receive rewards rather than equity, while equity crowdfunding gives investors company ownership and can require financial disclosure, compliance, and governance work.
- Kickstarter says it has added post-campaign services including shipping, taxes, and late pledges/e-commerce; Taylor also announced the Google-backed Next Wave Fund, offering $10,000 in non-dilutive capital to founders launching on Kickstarter.
- AMI Labs is still research-stage roughly six months after launch, with no near-term product plan. LeCun said it had 50–60 employees across Paris, New York, Montréal and Singapore and was speaking with potential customers and partners.
- Its JEPA/world-model approach aims to learn abstract representations of the real world and predict useful dynamics, rather than generate video pixel by pixel; the goal is to help systems anticipate the consequences of actions. LeCun sees industrial process control, predictive maintenance, anomaly detection and process optimization as near-term applications; the company expects research papers first, then B2B deployments through industrial partners, with consumer robotics and autonomous vehicles later. He gave no consumer-product timeline and said current robot makers still lack a way to make robots sufficiently intelligent.
- LeCun said the company has a long enough funding runway to take research risk but expects further fundraises because compute is expensive; it currently rents GPU-based AI-compute services and is too small to operate its own data centers. He said capital raised was 40% European, 33% American and 27% Asian/Middle Eastern, and that the company does not intend to sell.
- Barrett Lion is building Docset as a consumer-oriented alternative network intended to protect users from advertisers, trackers, malware and foreign adversaries; he connects the idea to privacy erosion, censorship, limited new internet protocol design and carrier-grade NAT, and says mesh and peer-to-peer networking are part of the approach.
- Lion says Docset has a mesh network with 196 domains that do not exist on the public internet, where users can register domains and lease IP or mesh addresses. He positions it beyond conventional VPNs’ leased entry-and-exit servers: its encrypted peer-to-peer apps support calls, video, chat and direct file sharing, with no intermediary server for transfers; he claims transfers up to 20 GB and post-quantum encryption.
- Lion says the company operates its own full stack from software to wire, including its AI and routing components; he reports 26 sites worldwide, 12 people involved and AI handling network operations. The company uses a US entity and a separate Swiss company for European operations. Lion’s prior infrastructure experience includes building a DDoS-protection service and a video-focused CDN, which he says he is now applying to consumer networking.
- Altman cited a steep math-capability trajectory: he said OpenAI models progressed from grade-school math three summers earlier to gold-level results at a leading international math competition last summer, and that one model had recently solved a Millennium Prize problem involving the Navier–Stokes equations.
- He warned that increasingly capable, autonomous systems could outrun institutions and human intervention, with recursive self-improvement potentially accelerating progress. Altman said OpenAI had slowed down unilaterally before and would do so again, and argued against training systems without strong alignment, monitoring, safety, and human-control guarantees.
- Altman called for democratically accountable government input and complementary national and international frontier-AI standards covering capabilities, risks, safeguards, human oversight, compliance evidence, incident reporting, and threat sharing. He said standards should not lock in incumbents or favor a business model, and should support both open and closed developers, new entrants, and established labs.
- OpenShell is an Apache 2.0, agnostic agent runtime designed to enforce governance outside an agent’s process so the agent cannot bypass controls while retaining flexibility. Sentry and BlueField are described as reference-stack components, not dependencies, in a broader application/runtime/infrastructure approach.
- The speaker reports 100+ partners, most already contributing to OpenShell, with no commitment required; the project is community-led and is on a path toward foundation donation. The speaker also reports partners using it for long-running agent fleets and extending it to robotics and physical AI, alongside integrations by platform and cloud providers and Microsoft in WSL.
- Agent-native identity and permissions remain unresolved, and partners are seeking solutions for the complexity of long-horizon, ephemeral agents. Policy creation, verification, attestation, trust, and application-layer identity, privacy, security tooling, and observability are also identified as areas needing work.
- LeCun argues that AI-mediated digital services will need diverse systems tailored to different languages, cultures, value systems, and interests, rather than being controlled by a small number of companies or one country. He sees open-source AI as a way to support that diversity, citing infrastructure software’s security, customizability, and ecosystem benefits; he says more-open LLMs are already being adapted for business verticals and for particular languages and cultures.
- LeCun says human- and animal-level intelligence and learning in AI are not imminent and would require scientific breakthroughs. He points to gaps in physical-world understanding and persistent memory, and says current LLMs do not reason or plan to the level desired.
- Kickstarter CEO Everett Taylor says AI-powered hardware is an “exploding” category on the platform, with projects raising millions; he also sees AI helping more people start projects, alongside laid-off tech workers turning to entrepreneurship.
- Kickstarter’s reward-based crowdfunding is non-dilutive and, Taylor says, best suited to physical products and hardware. Founders need a working prototype; a campaign can provide early customers, feedback, product-demand proof and leverage in a later VC raise. He recommends testing for product-market fit and preparing marketing, manufacturing and fulfillment before launch.
- Taylor says Kickstarter’s design and technology category grew 70% last year, and that some campaign-backed businesses scale to multi-million-dollar companies while others become profitable without needing venture capital. Kickstarter and Google also launched a fund offering $10,000 in non-dilutive support to founders launching on the platform.
Garry Tan endorsed CFO.ai as a finance agent for understanding company spending and whether it drives ROI, saying its out-of-the-box integrations enable near-instant setup. The linked testimonial quotes Jake Petoskey, CEO of fractional-CFO group Lighthouse Partners, saying he was impressed within an hour of setting up “Ari” and wanted an immediate call.
- Google DeepMind introduced Gemini 4 Argon for coding, enterprise knowledge work, and cyber defense, but initially limited access to Fairwind government users and trusted cyber defenders while refining guardrails. Google claims 13 wins across 19 benchmarks; Artificial Analysis scored it level with GPT-6 Astra at 53, and estimated discounted cost per task at $1.99 versus Astra’s $3.26. Argon used 62K output tokens per task versus Astra’s 27K, so the reported cost advantage came from pricing rather than efficiency. Google cites 1M output tokens, but Vals lists a 262K maximum; Artificial Analysis reached 1M through Long Decode Continuation, which pauses and resumes responses across calls.
- Anthropic reports open-weight GLM-5.3 achieved 50/410 end-to-end V8 exploits on ExploitBench, near Claude Mythos Preview’s 56/410, and full control-flow hijacks on 4% of its internal binary-exploitation tasks. The recap describes the model as widely downloadable, relatively cheap, and weakly refusal-tuned; commenters disputed Anthropic’s framing and cited defensive-security uses. Separately, merged llama.cpp support enables local inference for the 320B hybrid text-and-vision GLM-5.3-Flash, though a model-type naming mismatch may prevent some existing Unsloth quants from loading in mainline.
- The newsletter reports that OpenAI is near $70B in annualized revenue and is discussing a $30B raise at a $1.4T valuation, with its IPO pushed to next year. Flow, which builds AI tooling for hardware engineering, raised a $50M Series B at a $750M valuation.
- Perplexity’s contextual embedding model is open on Hugging Face; the newsletter reports a new state of the art on ConTEB and a 14.4-point lead over voyage-context-4 in answer recall@10 on turbopuffer’s private context benchmark, using 1 KB int8 vectors versus 8 KB. Cloudflare reported 648 ms p50 time-to-interactive for its rebuilt agent containers, described as 6x faster, with snapshots in beta; its model router reduced spend by about 30% in internal tests.
- Runway released an open-weight world-action model and says robotics-policy performance scales predictably with third-person video.
Sam Altman described 6.1 Sol as the fastest-growing model ever, while noting it had been somewhat slow under load and should now perform much better.
Sam Altman said users should be able to use their AI subscription wherever they need, linking to a post that identifies Sign in with ChatGPT (SIWC) and says its author would explain why it was shipped and address questions in the coming days.
YC added Vivian Shen and Raphael Schaad as full-time General Partners, bringing its partnership to 19; Garry Tan describes it as the largest in history and says all its members are YC alumni and founders. Shen co-founded Juni Learning (W18) and later founded AI college-prep tutor Acely; Schaad founded Cron (W20), which became Notion Calendar. Both had returned to YC as Visiting Partners after building products used by millions.
Garry Tan highlighted a video of a real-time generated agent and linked it to Tavus. Tavus says its Griffin model aims to reproduce human conversational cues—including expressions, movement, and turn-taking—and reports that it initially looked like a normal video call to the person watching; this is a qualitative, company-side assessment.
At DevDay, Sam Altman said builder energy was “crazy” and that people were making whole startups in a day. He also said Sign In With ChatGPT and Plugin Extensions have more potential than people realize.
Modus built an open-source Financial Audit Bench to test whether AI can do actual junior-auditor work, rather than only answer accounting questions; ModusAudit Labs announced the benchmark’s release. Garry Tan predicts that every job will eventually have evaluation benchmarks like this.
Flexport’s announcement says its MCP gives AI tools the ability to ship containers between any two places by any transport mode, including compliant customs clearance; Paul Graham amplified it as “AIs can now ship containers.”
- a16z says two AI companies have added more revenue this year than all public software combined, and frames the technology buildout as having just passed railroads as a share of GDP; it argues factories and datacenters are themselves products.
- a16z reports median AI-vendor spending in the top 1% of companies is 8× that of the top 10%, while only about 30% of S&P 500 companies report quantified AI impact. It identifies connecting models to company data and workflows—and broader enterprise diffusion—as a major opportunity over the next five years.
- a16z describes agent usage as still only a few million users, but sees potential to reach billions of internet users and an opportunity it says exceeds META/GOOG’s current $200+ annual monetization per U.S. user. Its five-year themes also include consumer agents, robotics, autonomy, AI x bio, personal health, enterprise diffusion, and American Dynamism.
Scott Kupor agreed with Arthur Macwaters’s forecast that medical records such as x-rays, lab reports, and blood panels will be put into LLMs, including by patients using their own models. Kupor cautioned that the standard of care will take time to shift, though he expects it to do so.
Y Combinator announced Startup School in Taipei for November 12, with talks and sessions alongside YC partners for builders from across Asia; featured founders include Andy Fang of DoorDash, Jeff Chang of Vest, Moses Lo of Xendit, and @mchiang0610 of Ollama. RSVP: https://events.ycombinator.com/startup-school-taipei