We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
1. Funding & Deals
Software M&A is active by count but thin underneath. Kroll’s Summer 2026 update, covering software M&A and public comps through June 30, annualizes 2,672 transactions and $240 billion of announced deal value; SpaceX’s $60 billion Cursor acquisition supplies roughly half, and excluding it brings annualized value closer to $120 billion—one of the lowest totals on record.
Strategic buyers are supplying the meaningful bid. Strategic acquisitions carried a 6.0x median EV/LTM-revenue multiple versus 3.3x EV/NTM revenue for public B2B software, and strategics represented 74% of software transactions. Kroll attributes that premium to corporates buying AI capability, proprietary data, and workflow position because they cannot build those assets quickly enough internally. For Seed-to-Series-A investors, the exit test is therefore less “can this scale?” than “does this team own a specific capability or workflow a strategic buyer will need?”
Scale alone is not earning the premium. Companies below $100 million in revenue ranged from 11.5x to 22.9x on precedent EBITDA multiples, and Kroll says the premium is moving toward smaller companies with specific, defensible positions.
2. Emerging Teams
A model-independent verification layer is the strongest early technical-team signal. An early-stage builder is separating claim generation from claim verification for financial use cases: an LLM produces a candidate claim, which is normalized, bound to evidence and assumptions, checked against constraints, subjected to proof or derivation and contradiction analysis, and returned as an auditable outcome. In a reported 66-case benchmark, structured fixtures passed 66/66, while live GPT-5.1-generated claims passed end-to-end only 19/66; the failures were concentrated in pipeline execution, claim binding, and contradiction detection, while the deterministic verification components continued to pass their tests. The assurance-layer thesis fits finance, risk, audit, and compliance, but the benchmark is internal rather than third-party validation and the trust model still needs empirical testing; the team is seeking researchers, engineers, and industry partners.
Cyborb is pushing AI app builders from code generation toward local execution and deployment. Its desktop agent plans a project, writes and tests code, generates images or audio, and publishes a live site from a plain-language request; it edits the user’s local files, is in beta, and shipped five releases in two days to fix issues encountered by users. The open question is differentiation: feedback places it against Lovable, Bolt, Replit, and v0, with the local workflow and direct publishing interesting but not yet a clear switching reason.
Huntme AI OSINT is a small but concrete traction signal. Its 18-year-old founder reports approaching 1,000 users and about ₹7.3K in revenue, with the current site built in two hours using 21 Dev and Antigravity. Data aggregation is outsourced to a friend for 20% of revenue; moving to a roughly ₹350/month self-managed server could improve control, but also makes the data pipeline an immediate scaling and relationship-risk diligence point.
3. AI & Tech Breakthroughs
Structured agent state is emerging as an alternative to ever-longer histories. Google’s SKILL.state proposal replaces full conversation replay with a structured current-state representation plus the latest observation, keeping input size roughly constant as a session grows. In a 100-step Gemini-3-Flash benchmark, it reported 0.94 accuracy using 65,000 tokens versus 0.91 using 1.1 million tokens for a LangGraph-style baseline—roughly a 94% token reduction. The failure mode is important: if the agent cannot anticipate what it will need later, it may omit that information from state and need to retrieve it again. A useful diligence question is whether systems measure recovery after a bad state write, not just average task accuracy.
Anthropic’s Model Hardware Standard moves agent infrastructure into the physical world. The research preview is a shared, model-agnostic specification for agents to operate multiple lab and manufacturing instruments in parallel; Anthropic says it can reduce bespoke hardware integration from weeks or months to hours or minutes and support real-time parameter changes and, in some cases, recovery from hardware errors. Its standardized driver exposes read/write primitives, makes devices discoverable, and creates a reference file containing device characteristics, adjustable functions, and enforced safety limits; agents can then sequence and monitor multi-device workflows through MCP, a CLI, or code files. This is still a research preview: Claude’s physical reasoning requires expert oversight, non-programmable hardware is not yet supported, and Anthropic is using the preview to develop additional safety evaluations before open-sourcing the standard.
4. Market Signals
Open-weight models are becoming the volume layer while fine-tuning closes task-level performance gaps. Exponential View reports that open-weight token share at Vercel reached 62% in a single day, up from 28% two months earlier. It also cites Bridgewater and Thinking Machines fine-tuning an open Qwen model to produce roughly 30% fewer errors than the best closed model on internal information-filtering tasks at one-fourteenth the inference cost; a 27B model can fit on a desktop Mac. This does not erase frontier-model value: service guarantees, harness quality, and reliability still differentiate Anthropic and OpenAI, and open models still generate revenue for inference providers.
Specialized compute is becoming a parallel market rather than a single Nvidia standard. The same essay reports that OpenAI’s Jalapeño chip used its own models to write kernels, cut roughly 10% from a major compute block, reached tape-out about 16 months after the first hire, and delivered a reported 1.5–1.9x Nvidia’s tokens per megawatt at peak throughput. It argues that differing requirements for latency, power, training, and inference create room for specialist chip firms.
The current-period follow-on to the Cursor dispute is architectural. OpenAI’s proposed withdrawal of direct model access after Cursor’s SpaceX acquisition made provider dependence visible; Harrison Chase’s response turns it into a design rule: separate the model from the harness and do not get locked in. A contemporaneous developer example reportedly preferred staying with OpenAI models and moving to Codex rather than staying with Cursor and switching models, showing that model preference can directly affect application selection.
AI infrastructure is also acquiring a labor and siting politics problem. A monitored post reports $50 billion of data-center construction this year, with construction unions partnering with OpenAI and Microsoft while nurses, flight attendants, and a university faculty union support moratorium efforts. It argues that construction unions’ roughly 11% membership, versus under 6% elsewhere in private-sector work, gives them leverage in local siting votes. Treat this as a directional social-license signal, but it belongs in infrastructure underwriting alongside power and permitting.
5. Worth Your Time
- Watch — 20VC: Should American Enterprises Work With Open-Source Chinese Models? | Only 10% of Neo-labs survive. Use the sections on outcome-based pricing and verifiability, then the discussion of stateful harnesses, open-model specialization, and why most generic neolabs may not remain independent businesses.
Read — 🔮 Unbounded self-improvement and its limits #599. The essay provides the clearest current pairing of enterprise open-model adoption, fine-tuning economics, and specialist-chip opportunity.
Read — Anthropic’s Model Hardware Standard research preview. The useful material is the standardized device driver, machine-readable safety limits, and the gap between autonomous orchestration and the expert oversight still required in physical environments.
Thread — Harrison Chase on separating model and harness. A concise framing of why portability is becoming an application-architecture requirement as model providers compete with the products built on top of them.
Direct answer: Anthropic presents the Model Hardware Standard (MHS) as a research-preview specification for enabling AI agents to safely operate multiple physical devices through a common interface; it is intended for scientific labs, advanced manufacturers, and other settings using hardware with programmable interfaces.
What it does: MHS supplies a standardized driver that translates between an operating system and hardware, exposes common
read/writeprimitives, makes devices discoverable, and generates a device reference file from natural-language hardware tags. That file records device characteristics, adjustable functions, and safety limits to be enforced. Agents can control devices through MCP, a command-line interface, or code-file APIs, then sequence, monitor, and adjust multi-device workflows in real time.What it connects to: Anthropic says MHS works with any device that has a programmable interface. Named examples include microscopes, liquid handlers, robotic arms, laser and camera systems, nucleic-acid purification equipment, and quantum-computing hardware; partner projects include laser-scanning microscopes, Tecan Fluent liquid handlers, Doosan and Universal Robots arms, and QIAGEN’s QIAsymphony Connect platform.
Operational autonomy: The standard supports orchestration across instruments, live parameter changes, long-running command chains, and in some cases recovery from hardware errors without intervention. Anthropic describes Claude using camera feedback to adjust and align a laser, then packaging the learned procedure into a deterministic script.
Access/permission controls described: The current release is a research preview offered first to a group of scientific research labs and advanced manufacturers; prospective participants apply through the MHS site, and Anthropic is inviting stakeholders to join a waitlist. Anthropic says it plans to open-source the standard after further work.
Safety controls and safeguards: The concrete technical control Anthropic specifies is the device reference file’s enforced safety limits. Anthropic also says Claude’s physical and spatial-reasoning limitations require expert oversight, illustrated by researchers having to correct Claude’s interpretation of foaming as a physical sample failure rather than a software bug. The research preview is intended to support additional safety evaluations, stronger protections for physical-world use, a physical safety roadmap, and expanded safeguards-policy and enforcement coverage; planned open-source guidance will include research-preview findings.
Important limitation: MHS does not yet support hardware without a programming interface, and the announcement describes access gating, enforced device safety limits, and expert oversight rather than a detailed end-user authorization scheme or per-action approval workflow.
- Model economics and specialization: Factory CTO/co-founder Eno Reyes argues that AI economics should be judged by outcome rather than token input: a high-quality model that completes a code review immediately can be cheaper than a low-cost model consuming 50 million tokens, so the smartest model may be the cheapest system. He expects commodity task execution to be dominated by open models, with enterprises using post-training to create private specialized models between generic commodity and expensive frontier models.
- Technical frontier and infrastructure: Reyes calls verifiability the most important property of current AI systems and identifies models that can create verification where none exists as a path into harder legal and healthcare work; he cautions that poorly designed evaluation criteria can create the wrong incentives. For agentic workflows, he argues that simple gateway routing is insufficient: systems need stateful, in-task allocation of intelligence, with the harness becoming the application layer that maintains logic and state. He says continual learning is accumulating at the harness layer, making ownership of data, outcomes, and workflows—and the option to run on-premises—strategically important for enterprises.
- Open-model investment split: Reyes forecasts that 99% of workflows will run on open models within three years, while the remaining 1% of tasks could represent 30–40% of the future economic value of intelligence; he sees frontier differentiation concentrated in areas such as bio research, advanced AI development, security, and defense. This implies a low-cost, high-volume open-model layer alongside a small, high-value frontier layer.
- Portfolio filter and competitive risk: Reyes predicts that 80–90% of neolabs could cease to make sense as independent businesses within 18 months. He favors companies attached to durable, proprietary workflows that remain valuable as frontier models improve and serve persistent domains such as legal, while generic intermediate knowledge-work tools around Excel and Jira appear vulnerable. More broadly, he argues that technology itself is no longer the moat; usage data, workflow ownership, and consensus systems of record are more durable, while AI-native development could eventually disrupt agile software processes.
- Company and talent signals: Reyes says Factory Private is among Factory’s most popular offerings and that tens of thousands of self-service users use its product daily, while the company deliberately avoids broad consumer subsidies and emphasizes enterprise outcomes. Factory’s stated hiring plan is to source 100% of future hires through acquisitions of companies and their founders or teams, prioritizing builders with direct problem alignment, demonstrated execution, and user adoption over conventional pedigree.
- a16z announced the Machine Age Fund, a new $1.1 billion vehicle backing founders rebuilding the AI physical and infrastructure stack—chips, memory, networking, systems software, power, and machines that bring AI into the physical world.
- The fund’s thesis is that model progress is outpacing the memory, interconnect, power, and cooling beneath models, making these infrastructure constraints company-building opportunities and pulling strong teams from pure software into complex hardware.
- Ben Horowitz and Martin Casado argue that the hardware wave will favor more experienced founders because building a company while learning an unfamiliar product domain is unusually difficult; they also expect new hardware companies to create a generation of founders, analogous to the entrepreneurial spillover from SpaceX.
- Anthropic’s MHS is described as a new hardware standard that lets AI agents directly control printers, industrial equipment, and operational systems. The discussion identifies a 50 ms pre-actuation window as the practical intervention point for behavioral drift and says the industry has not settled what controls should operate within it.
- RuntimeAI describes inline runtime-policy enforcement that evaluates every outbound action before dispatch and blocks commands when an agent has drifted from its sanctioned scope, including within the 50 ms window. A reply flags single-policy-engine failure risk and asks how policy updates apply while an agent is running.
- Founding-team signal: Harry Stebbings says he sought to invest after meeting MatanSF and highlights co-founder Eno Reyes as a highly capable technologist with a strong understanding of building high-margin, efficient AI businesses.
- AI evaluation frontier: The post argues that verification systems for domains without benchmarks—systematically defining what “good” looks like—could make complex, high-friction human tasks reliably automatable.
- Infrastructure and application architecture: Frontier model providers are framed as choosing between high-volume inference and high-margin applications, while model-locked applications may conflict with enterprise demand for the best result across multiple models. For agentic performance, the post argues that simple gateway routing is insufficient; systems need to operate statefully inside workflows using context, execution history, and next actions.
- Defensibility and enterprise sovereignty: The post predicts that 80% of “Neo Labs” will die and says winners will anchor to durable enterprise workflows that remain valuable as frontier models improve. It identifies the harness layer—not closed model APIs—as the locus of continuous learning and workflow optimization, making owned harnesses and learning loops strategically important.
- Open-source model investment signal: Chinese open-source models should, in the post’s view, be assessed by technical security rather than origin alone; despite possible creator bias, improving open weights could power a growing share of standard enterprise workflows.
- Open-weight and fine-tuned models are gaining enterprise traction: Vercel’s open-weight token share reached a single-day record of 62%, up from 28% two months earlier, while some Western firms are moving workloads to Chinese open weights and Thomson Reuters built an in-house Qwen model to reduce costs. Bridgewater and Thinking Machines fine-tuned an open Qwen model with expert-labeled data, achieving roughly 30% fewer errors than the best closed model on internal information-filtering tasks at one-fourteenth the inference cost; Trainloop reports that its small Qwen 3.7-27B model can outperform GPT 5.6 Sol on specific fine-tuned tasks at a fraction of the cost and fit on a desktop Mac.
- The open-model shift does not eliminate frontier-model differentiation: service guarantees, harness quality, and reliability continue to drive business for Anthropic and OpenAI, while the revenue impact of open models remains uncertain because inference providers still get paid.
- AI is opening a specialized-chip opportunity: OpenAI’s Jalapeño chip used its own models to help write kernels, cut roughly 10% from a major compute block, reached tape-out about 16 months after the first hire, and reportedly delivered 1.5–1.9x Nvidia’s tokens per megawatt at peak throughput. The resulting market is expected to become more heterogeneous across latency, power, training, and inference needs; Cerebras powers ChatGPT’s fast-response mode, while Fractile has a deal with Anthropic for low-latency inference chips.
- Model-adjacent tooling can materially change outcomes: SwarmOS reportedly raised GPT-5.6 Sol’s ARC-AGI-3 Public score from 13.3% to 100%, reinforcing the investment case for harnesses and orchestration layers alongside base models.
Novel interactive-media paradigm: An unidentified creator built an “Interdimensional Cable” connected to a Rumble chat, where random prompts drive continuous generation and real-time scene changes—described as an “actual Infinite Jest.” Sriram Krishnan endorsed the demonstration as “bad ass.”
- Technical founder/product: An Iranian founder learned from a friend focused on machine learning for finance, tested major image-generation models, and fine-tuned Stable Diffusion using RTX 2080 Ti and 3090 Ti compute. The resulting GitHub release, Open Journey, was positioned as a “Midjourney quality” Stable Diffusion 1.5 fine-tune.
- Organic distribution and early traction: A Hacker News post triggered a large wave of LinkedIn interest and German media interviews. The founder then built the Mann-E platform and collected more than 180 tester email addresses in three days, against an expectation of 10–20. Media coverage was also generated inbound, without paid promotion.
- Monetization signal: Mann-E was marketed as a cheaper Midjourney alternative; the founder says news-site referrals brought users who found roughly comparable quality at about one-tenth the price, producing strong conversion. The founder reports receiving grants and investment and making substantial sales, but provides no financing terms or investor identities.
- Cautionary market signal: Mann-E was later shut down and is currently dormant as a blog. An internet blackout reduced prospective Iranian customers’ ability to afford connectivity configurations, while newer AI releases made the product irrelevant.
- Immigration and technical-talent policy signal: Jason Calacanis argues that the U.S. should secure its border while actively recruiting elite performers from countries including China, Russia, and Iran, framing that talent migration as a gain for the U.S. and human freedom. He supports a large H-1B fee and says adding $20,000–$100,000 could deter employers from using lower-paid foreign IT workers to replace U.S. workers; this is his stated position and rationale, not independently verified policy analysis. For startups dependent on international technical hiring, the post flags potential regulatory and recruiting-cost risk.
- a16z announced the $1.1 billion Machine Age Fund to back founders rebuilding the AI infrastructure and physical stack, including chips, memory, networking, systems software, power, and machines that bring AI into the physical world.
- The fund’s thesis is that model progress is outpacing the underlying memory, interconnect, power, and cooling infrastructure, creating company-building opportunities as strong software teams move into complex hardware.
- a16z argues that scaled incumbents such as Nvidia may leave substantial room for startups applying first-principles innovation to under-served infrastructure problems, because incumbents prioritize their largest existing growth opportunities.
- OpenRouter reported that heavily discounting GPT 5.6 Terra and Luna drove a 13.8× increase in token usage, an example of the Jevons paradox in AI inference: lower costs can expand total consumption.
- Jason identifies cheap or unmetered tokens—particularly through local models and on-premises servers—as a major driver of open-source and local-hardware trends, suggesting that falling inference costs may unlock substantially more usage rather than simply reduce spending.
Jason sees robot-operated, 24-hour cafes serving exceptional beverages as an emerging market trend and says Cafe X was ahead of its time.
Runway’s HDR ProRes Conversion was described as “Bonkers,” while a related post claimed: “The quality you need to make everything you want.”
- OpenAI proposed removing its model access from Cursor on November 12, triggering public responses from Cursor’s CEO and Anthropic’s founder. Allie K. Miller frames the dispute as a trust and platform-dependence issue: OpenAI appears unwilling to continue supplying a Musk-led/SpaceX-linked company, while Cursor wants to contest the move and says OpenAI represents only a small share of its traffic; Anthropic reaffirmed its relationship with Cursor and emphasized its compute support.
- The episode is a competitive signal for AI developer tools: model preference can directly influence product selection, with one developer reportedly choosing to remain with OpenAI models and move to Codex rather than stay with Cursor and switch models.
- A forward-looking competitive thesis predicts that the OpenAI–Cursor conflict will accelerate Grok’s push to catch up with Astra and Fable, potentially with a significant price advantage.
- Legora shows a sharp reported adoption curve in legal workflows: one lawyer’s usage rose from roughly 10% of tasks six months earlier to 80% the following week, and the lawyer said they would be “SO SO upset” if it were removed. This is an anecdotal but strong product-market-traction signal, with the post arguing that legal work could follow coding’s adoption trajectory.
- Verifiable-reward reasoning: The timeline identifies 2024–2025 verifiable-reward reinforcement learning as a major direction, describing DeepSeek-R1 as demonstrating that checkable math and coding rewards can improve models without human-preference labels.
- Agentic AI architecture: Agent systems are framed as a 2025–2026 convergence of reasoning, tools, memory, and execution; ReAct’s interleaved reason–act loop is offered as the core architectural pattern, with the caveat that progress reflects many incremental contributions rather than one defining breakthrough.
- Scaling and compute/data strategy: Transformers and scaling laws are presented as the foundation for predictable gains from increasing parameters, data, and compute, while compute-optimal training requires scaling data alongside model size.
Bindu Reddy predicts that OpenAI’s Astra will likely launch Thursday and be more capable than Anthropic’s Fable. She attributes to Astra persistent performance on very difficult problems, coordination of thousands of agents at scale, system-level rather than task-level reasoning, and a highly creative, action-oriented behavior profile.
- Early-stage technical founder signal: A third-year diploma computer-science student built and deployed an end-to-end deepfake-audio detector as their first substantial machine-learning project.
- Technical approach and limitation: The prototype uses EfficientNet-B0 on mel spectrograms trained with the ASVspoof 2019 LA dataset; it reports 0.88 F1, 0.99 precision, and 0.79 recall. The builder identifies recall as the weak point, saying a 0.4 decision threshold did not help much and that the model may miss attack patterns absent from training.
- Productization and explainability: The deployed system adds Grad-CAM spectrogram visualizations and Groq-generated plain-English explanations, supports file upload and live recording, and takes roughly 6–7 seconds per prediction including inference and explanation generation. It runs on FastAPI/Railway with a Streamlit Cloud frontend.
- Bindu Reddy predicts that OpenAI’s “Astra” will launch in September and become a major development; she expects Cursor customers to want access and be dissatisfied if it is unavailable, signaling potential competitive pressure on AI coding platforms.