We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
1. Funding & Deals
Ben Horowitz’s largest-ever check is backing Travis Kalanick’s industrial-AI vehicle. Kalanick defines the company as “systems of software, sensors, robotics and AI” for automating industrial sectors, mapping manufacturing, real estate, and transport/logistics onto physical-world equivalents of the CPU, storage, and network. Its food-infrastructure example targets robotic production and delivery at roughly $6–$8 per meal.
The disclosed financing signal is relative rather than a conventional price comp: Horowitz calls it his biggest check ever, while Kalanick says it is several orders of magnitude larger than Uber’s first $4 million pre-money round. The transaction also required merging separate entities with different investors. The team is recruiting from Uber’s operating bench, including former CFO Gautam, former engineering SVP Ganesh, and former Uber ATG lead Eric Myhofer. The diligence question is whether this is a scalable industrial platform or a collection of difficult vertical bets; the company’s own examples span food, mining, and specialized industrial transport.
2. Emerging Teams
Agent correctness is becoming an investable wedge separate from agent capability. An ex-PM for AI at a fintech describes a failure mode in which an agent changed loan records, returned “done,” and left a green trace even when writes had not reached the underlying system. He is building post-hoc verification against the real system rather than the agent’s own logs, but still lacks live production agents to test against and is seeking three to five design partners in billing, CRM, and internal operations. This is a particularly clear early-stage test: the product either catches silent state divergence on real traffic or it does not.
MARGINAL is taking a similar reliability-first approach for coding agents. The open-source governance layer watches for repeated actions, weak progress, redundant verification, and low-value continuation; it starts in Shadow Mode, earns permission to block only after proving reliability, and falls back when confidence degrades. Its proposed next layer is counterfactual evaluation: whether the agent would actually have done better without the intervention. The ensuing design discussion recommends paired governed-versus-ungoverned runs from the same state, repeated seeds, versioned policy and reason codes, and a regression corpus for false stops and missed loops. The investor takeaway is that governance overhead and task outcomes—not token savings alone—should determine whether these systems earn enforcement rights.
Valor Atomics is underwriting nuclear scale through iteration speed. The less-than-three-year-old company says its OR250 reactor reached first criticality in November and produced power “a few days” before the interview. Its internal metric is the time between atom-splitting events: two years and four months from incorporation to the first split, about seven months to the second, and an eventual goal of minutes. Founder Isaiah Taylor says the team deliberately recruits high-agency hardware builders, including people who entered nuclear without prior reactor experience.
The financing and deployment strategy is unusually execution-heavy for nuclear: Valor is using equity capital to put reactors on the ground before project finance becomes available, and proposes building gigawatt sites speculatively so customers can follow the power rather than determine the construction schedule. That makes the company worth tracking as a hard-tech formation signal, while keeping the commercial scale thesis separate from the demonstrated first reactor.
3. AI & Tech Breakthroughs
GLM-5.3 makes post-training, not just pretraining scale, the center of the open-model competition. Interconnects reports that Z.ai’s model is initially available through its coding plan, with API access and Hugging Face weights planned, and places it near the frontier of agentic coding benchmarks at roughly 750 billion parameters. The analysis says GLM-5.3 uses the same base model as GLM-5.2 with substantially extended post-training, and quotes Z.ai’s framing: “Scaling post-training is all we did.” The article also argues that Chinese labs’ faster public release cadence—days rather than the months typical of OpenAI or Anthropic—may compound if model-improvement loops depend on user data.
The release carries a concrete dual-use test. Z.ai describes GLM-5.3 as its most capable cybersecurity model and says it will stage access through security partners, then API availability, before releasing complete weights; the analysis notes that smaller, easier-to-modify open models may make any single lab’s safeguards insufficient.
Document AI is being evaluated against production degradation rather than clean PDFs. A current-period ExtractBench repost points to regulatory filings, hand-filled tax forms, photocopies, and sensor noise; its summary says Codex performs well on scans but poorly on rotated documents, while OCR systems show the inverse pattern. This is also a benchmark-governance signal: Vik Paruchuri said fixing scoring bugs moved Datalab from 65% to 93.6%, and LlamaIndex’s Jerry Liu said the team would incorporate the configuration changes and update the results. For diligence, reproducibility and failure-case coverage deserve the same scrutiny as the headline score.
AI-biology is moving from sequence prediction toward wet-lab design. A Lightspeed interview describes Stanford and ARC Institute researchers using a genomic model trained on roughly a couple million bacteriophage genomes to design new bacteriophages targeting E. coli; the speaker presents the resulting variants as a first-of-its-kind result and flags the biosecurity implications. The important qualification is that the model produced a design roadmap: human scientists synthesized the viruses in a wet lab, and manufacturing and regulatory controls remain meaningful bottlenecks.
Valor’s physical proof is more useful than its long-range projections. The company says it connected an Nvidia Blackwell system directly to the reactor and hosted a website from that chip, while planning a passive-safety demonstration that shuts off electrical supply and active cooling so natural circulation removes decay heat. Those are company-reported demonstrations, but they give investors more to underwrite than a reactor design deck alone.
4. Market Signals
Compute is becoming a longer-lived asset and a routing problem. CoreWeave says it remains sold out of prior-generation GPUs as well as current SKUs and has signed an A100 contract extending into 2029, despite the chip having launched in 2020. Clouded Judgement interprets the evidence as support for useful lives of nine years or more and points to a software layer that routes each inference request to the right chip as workloads spread from frontier-complex to trivial. The opportunity is therefore not only new capacity; it is asset utilization, workload routing, and economics across heterogeneous fleets.
AI coding is consolidating around compute, product data, and the outer loop. The Cursor/SpaceX account describes a $60 billion all-stock acquisition that gives Cursor “an order of magnitude more compute” for post-training and product-linked experimentation. This is a late-stage strategic transaction, not an early-stage financing comp, but it is directly relevant to AI competition: Cursor says agents have pushed productivity gains beyond tab completion, while the resulting code-review bottleneck has created demand for tools around testing, integration, and verification. Its revenue mix had shifted to almost 80% B2B, with large enterprise deployments expanding rapidly.
Agent products are being differentiated by domain feedback loops and APIs. Klaviyo reports that its Composer marketing agent reached 95,000 users in its first month, with about one-quarter returning weekly, and that the first prototype was built over a weekend by other agents. Its claimed moat is not the base model: Composer receives a proprietary real-time feed of consumer responses and uses a coaching agent to score proposed campaigns for engagement and revenue. The company is also treating every employee as an L3 agent operator and pushing toward headless, agent-friendly APIs rather than human-only interfaces.
The founder pipeline is becoming more singleton-heavy. Paul Graham reports that single-founder companies rose from 9% to 18% in recent YC batches, attributing the increase—tentatively—to AI enabling one person to get more done. He still argues that a cofounder’s value is stress-bearing, not just additional output. That is a useful counterweight to claims that agent leverage removes the need for complementary founding teams.
5. Worth Your Time
- Watch — Travis Kalanick on Building Atoms After Uber. Focus on the physical-world CPU/storage/network analogy and the definition of industrial AI; it is the clearest expression of the period’s industrial-execution thesis.
- Watch — AI Creates New Viruses, Building AI for Healthcare & Moats in Healthcare IT. The useful segment separates model-assisted biological design from wet-lab production and lays out the associated biosafety question.
Read — GLM-5.3: How Chinese labs keep stride with the frontier. Read it for the post-training thesis, release-cadence comparison, and dual-use caveat.
Read — ExtractBench. The current-period repost links the paper and site and gives a compact picture of why real document conditions defeat one-size-fits-all extraction claims.
