AI Startups Raise Billions, Hugging Face Passes 3M Models, and Perplexity Expands Agent Controls

Key Takeaways

1. An AI chip startup raised $700M at a $21B valuation, delivered its first rack to Jane Street, and counts Jane Street, Kleiner Perkins, Sequoia, A16Z, Peter Thiel, BCV, and Blackstone among investors. — via 1
2. Menlo-backed Wispr Flow raised $280M at a $2B valuation with roughly 40x YoY growth, while Higgsfield raised $400M at a $5.4B valuation and grew ARR from $20M to $700M in 12 months. — via 1
3. Hugging Face Hub crossed 3 million models and published ICML reproduction results involving 1,221 humans and coding agents on 2,226 papers. — via 1 2
4. Perplexity expanded Computer agent controls with connector permissions and email tasks, and says River API outperformed Tinker in RL tests. — via 1 2 3
5. NVIDIA launched TensorRT Model Connect for two-command conversion of Hugging Face models, built using an OpenAI Codex agent, plus Context-Matched Distillation for video generation. — via 1 2
6. Ethan Mollick cited early evidence that AI is accelerating science discovery in cybersecurity and math, but not yet algorithms. — via 1

1. Funding and Hardware Milestones

1. An unnamed AI chip startup announced $700M in funding at a $21B valuation and said it has already delivered its first rack to investor Jane Street. The investor list also includes Kleiner Perkins, Sequoia, A16Z, Peter Thiel, BCV, and Blackstone, making the round a strong signal for AI-native hardware demand. — via 1
2. Menlo Ventures portfolio companies are raising at scale: Wispr Flow raised $280M at a $2B valuation (Menlo led) with about 40x year-over-year growth and over 60 billion dictated words; Higgsfield raised $400M at a $5.4B valuation (DST led) with annualized revenue leaping from $20M to $700M in 12 months. The numbers show application-layer AI companies can move from small ARR to massive scale quickly. — via 1
3. Deedy argues that despite these growth numbers, app-layer winners are hard to build: software is easier to copy than ever, so founders need extreme focus on product, growth, retention, monetization, infrastructure, enterprise sales, and team building. — via 1

2. AI Models and Open Source

1. Hugging Face Hub passed 3 million models, which the team frames as accelerating momentum toward open, distributed AI. — via 1
2. Hugging Face's ICML paper reproduction challenge produced 6,816 reproduction logs, 2,962 cloud tasks, and 35,908 judgments from 1,221 humans and coding agents reproducing 2,226 papers. They noted the next million Hub users may be AI agents rather than humans. — via 1
3. Sentence Transformers v6.0 adds MultiVectorEncoder, making ColBERT-style late-interaction models a first-class model type alongside dense, sparse, and rerankers. — via 1
4. NVIDIA's new TensorRT Model Connect public preview converts Hugging Face models to end-to-end TensorRT inference in two commands, with no intermediate ONNX export, and allows native C++ API execution. The project itself was built by an OpenAI Codex agent with humans reviewing and guiding; it is open source. — via 1
5. NVIDIA also introduced Context-Matched Distillation (CMD) for fast, controllable autoregressive video generation, enabling frame- and block-level generation and precise camera control. — via 1
6. Aravind Srinivas says the progress on dense models running locally is "incredible" and points to the near future of local AI, an observation consistent with the broader open-weights trend. — via 1

3. AI Agents and Tooling

1. Perplexity's Computer now supports connector permission controls (Allow / Always Ask / Deny) that let users approve an action once or keep tools available for a thread; loop runs follow thread approvals, and the controls are live for all Computer users on the web. It also added email tasks, where sending, forwarding, or CC'ing [email protected] creates a task that runs as a normal session with the same audit records. — via 1 2
2. Perplexity says its River API outperformed Tinker in reinforcement-learning tests using identical training code, after focused engineering on routing and replay. — via 1
3. Hamel Husain shared updated eval skills with a new error discovery skill that hands AI outputs or trace files to a coding agent, which builds a custom review app with smart sampling, classifies failure modes, and retrieves relevant examples; a new startup skill routes context to workflows for debugging or auditing eval pipelines. He argues the practical recipe is human experts + deterministic evaluators, with fully automated evals not yet feasible. — via 1 2
4. swyx says vibe coding can produce open-source SaaS alternatives: in the Kill My SaaS hackathon, 69 people completed submissions replacing Sessionboard, with 69 submissions. — via 1 2
5. swyx also recommends a free, open-source context engineering workshop covering chat history management and prompt caching, which he says can cut about 90% of token costs while keeping agents in long conversations. — via 1

4. Research Insights and Practical Observations

1. Ethan Mollick, drawing on collaborative work with Nate Rush, reports early evidence that AI accelerates scientific discovery in cybersecurity (sharply) and mathematics (some), with no clear acceleration yet in algorithms. — via 1
2. Mollick notes that LLMs having theory of mind is a major advance, but they still struggle when required to balance multiple audiences, such as separating end-user needs from creator perspectives when coding. — via 1
3. Mollick also argues AI labs should pay more attention to variance in creative tasks: getting varied creative output from intelligent models requires significant effort, which limits their usefulness in problem-solving and creative work. — via 1
4. swyx highlighted a historical, unverified claim from George Hotz that GPT-4 was a 220B x 8 mixture-of-experts model in June 2023; treat this as a claim rather than confirmed fact. — via 1