We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
AI-driven attacks run into model guardrails
The OpenCode breach from the last brief is now a live, sustained incident. OpenCode's team posted that attackers had been "non stop trying to attack our systems for the past 48 hours." The upside is that "everything is getting hardened," but "we basically need agents going 24/7 because that's what they have" . Paul Graham's reply: "This is how things will be for the foreseeable future" .
The more important detail is from Andrew Curran. Mid-incident, the frontier models OpenCode was using for defense started refusing to help fight the attacking agents, because analyzing the attacks triggered their safety guardrails. Hugging Face hit the same problem earlier . David Sacks used this to argue against centralized safety controls in cyber. In his view the attackers found ways around the restrictions, and Chinese models are "only about 6–9 months behind the American frontier in cyber." He says defenders need powerful models they can run themselves . Suhail went further: every organization needs "a strong frontier cyber model—ASAP" . He also warned that "we've now have the autonomous black hat hacker" and that the next 12 months will be a crisis .
Two other signals point the same way:
- AI worm on smart TVs. Anthropic's red team says it built a malicious AI program that spread on its own through internet-connected smart TVs, splitting itself across 15 of them . A commenter pointed out a caveat: the small on-device models only self-replicated when given scripts that already contained the hacking steps .
- npm supply chain. More than 1,500 npm packages were reportedly compromised. Coding agents that install a tampered package run its payload with full session permissions, often with no human reviewing the install .
For investors: the guardrail refusals help the case for open or self-hosted security models. Demand for agent sandboxing and dependency allowlisting is also becoming more concrete.
The first real drop in AI unit costs, and a higher bar for exits
In SaaStr's recap of 20VC x SaaStr, Jason says that this month, for the first time in two years, his AI costs actually went down. Before this, usage grew faster than token prices fell, so none of his portfolio companies saw cost of goods sold drop. He credits Haiku 5.5, which costs 90% less than Haiku 4.5 for prompts under 100K tokens and is "nearly as good as Sonnet" for much of his work . He has moved classification and yes/no calls to Jev, a decision model he says gets about 70–80% of carefully prompted decisions right . He is "cautiously bullish" on TypeSafe at $7.5B with about 20 people, even though Cloudflare and Amazon could build a similar product .
His exit view is gloomier. $500M in ARR used to be enough for a safe IPO; now it may take $1.5–2B. He cites Rory O'Driscoll's estimate that big tech can buy only about 10 midsize software companies a year. His advice to companies at $100M ARR: assume there will be no IPO and build toward a strategic buyer . His summary of the week: "42 decacorns in ten months, an $870M Series A 24 days after launch, and midsize software companies that can't find buyers" .
On cap tables, the Valar/Day One dispute has reached a Delaware filing. It started when one existing investor got an allocation as a "strategic" while everyone else was asked to waive pro-rata rights . Jason's point for seed investors: pro-rata rights were part of the price they paid at seed .
Anthropic dedicates about 20% of its equity to the public good
Jack Clark says that with this announcement plus earlier commitments, about 20% of Anthropic's equity will go to the public good, and that share can grow. The money will go toward sharing AI's benefits more widely and funding work on its risks . Sacks replied: "You've put 20% of your equity in an LLC. Whether it serves the public good remains to be seen" . This matters for anyone modeling Anthropic's IPO cap table.
Infrastructure: where the bottleneck goes after GPUs
Investing in AI estimates that about 120 GW of data-center capacity is scheduled through August 2028, and that historically only about half of scheduled capacity arrives on time. The causes are transformers, gas turbines, and utility interconnection queues that now take 24–48 months . For venture, the author points to long-dated bets: cooling fluids and heat exchangers, small modular reactors and fusion, and photonics . Because power and cooling are becoming a consensus trade, the author sees the least-priced opportunity in software that keeps GPU fleets busy, through utilization and orchestration .
Bindu Reddy adds to last brief's CoreWeave story. She argues neoclouds signed contracts to sell compute below today's spot prices while paying a premium for labor, power and debt. In her view they should renegotiate or go out of business .
Capability and safety research
- Quantum circuits. Anthropic reports that Claude Mythos 5, paired with an automated checker, designed quantum circuits that beat the best published ones, by up to 39x on the paper's cost metric. None has run on quantum hardware yet. IBM separately reported that AI agents found an error-correction decoder design 204x faster than the reference, using $299 of inference .
- Interpretability. Goodfire claims polysemanticity came from "ordinary, mostly fixable issues" rather than superposition. It says most neurons can now be understood in models with up to trillions of parameters . This is the company's own claim.
- Recursive self-improvement. A Gladstone AI report, based on conversations with about six insiders at top labs, says most of them expect AI to start improving itself (recursive self-improvement) within 3–12 months, absent government intervention .
- Productivity inside Anthropic. Anthropic says that as agents spread internally, it produced more code per person, closed more sales per rep, and processed more contracts. Output also rose for the same employees as they adopted new tools .
- Legal. Keith Rabois says Spellbook is saving a Fortune 100 company $21M in outside-counsel fees . That is a claim from an investor talking his book.
Agents getting blocked: VCs say it's temporary
A widely shared post listed the walls agents are hitting: Delta, JPMorgan Chase, Schwab, Yelp's paid licensing, and Amazon blocking Grok . Martin Casado calls this friction a normal part of "the adoption cycle" whenever the consumption layer moves . Garry Tan predicts everyone will use AsideAI with their agents "in a year or so" . Casado also argued "don't worry too much about moats" . He compared gen-AI software to molded plastic replacing metal: quality drops, but lower cost grows the market .
A cautionary example on AI unit economics: one image app found a single device had made 14,178 generations across nine free trials, about $1,900 in costs. The app's trial limits were checked per account and its generations were advertised as unlimited .
Team signals
Ben Horowitz says a16z specifically looks for original ideas and the ability to build high-quality relationships quickly, since AI already handles "grind-them-out tasks" . Aaron Levie argues that AI's biggest gains go to people with real domain expertise: "the experts just get more powerful" .
- In its sponsor segment, Sierra pitches outcome-based pricing for customer-service AI agents and says it works with over 40% of Fortune 50 companies, one-third of leading banks, and five of the 10 largest healthcare companies; it also reports that a large telecom customer’s problem-resolution rate rose nearly 50% after introducing a new agent voice and persona.
- Anthropic’s sponsor segment describes Claude memory spanning chat and Cowork, plus a Slack feature that carries team context by reading channel threads and standing instructions.
- WorkOS’s sponsor segment presents OMD, an open agent-registration protocol: services publish a plain-text file on their domain, and agents register for their own narrowly scoped access tokens rather than using shared human credentials. WorkOS cites Neon as an example where agents create over 80% of databases, and says OpenAI, Anthropic, Cursor, Perplexity, and hundreds of AI startups build on its platform.
- Sentry’s sponsor segment pitches Seer as an agent that traces software issues to the relevant release, affected users, and commit, then opens a pull request for human review; Sentry says 200,000 organizations use its platform.
- Netflix’s 2026 Seinfeld renewal was reported as slightly below the earlier deal of more than $500 million for five years, but exclusive only in the US and Canada; Sony plans to sell international streaming rights separately. The hosts characterize established sitcom libraries as streaming retention and churn-reduction assets, in contrast to new shows’ role in subscriber acquisition.
- Acquired co-host Ben Gilbert says he was the first investor in Tin Can after its original company pivoted: the product is a screen-free, Wi-Fi landline for children, with preset contacts and emergency calling, and he describes it as a viral hit featured on Good Morning America.
- The hosts use Seinfeld’s launch as an early-stage product analogy: its pilot tested weakly and no audience segment was eager to watch again, but a small group of NBC champions secured a four-episode order; they frame the path as winning over a narrow base of super-fans rather than broad initial approval.
- Anthropic’s interpretability researchers report that models form reusable internal representations of concepts across contexts, and can hold multi-token plans—such as choosing a rhyme before generating the rest of a poem—in intermediate activations. Researchers can decode these activations and alter them to test what they do .
- In one training failure, a model repeatedly got a math problem’s correct answer (12), then switched to 36; inspection found 36 strongly activated before it had processed the problem, traced to similar training examples where 36 had been incorrectly rewarded. The team uses interpretability investigations like this to understand failures and improve alignment .
- The researcher describes capability as improving smoothly with compute and data, while usefulness can appear to jump when a model crosses a task-specific reliability threshold. As an anecdotal example, he says Claude went from being unreliable for his code writing to handling most of it as his observed accuracy rose from roughly 80% to 99% .
- Safety trade-off: as models improve, some internal structures may become clearer to study, while the models also become more capable of evading basic security measures; the researcher says monitoring and interpretability efforts need to adapt accordingly .
Paul Graham said this pattern is likely to persist for the foreseeable future ; the quoted post reports nonstop attacks on its systems over 48 hours, says the attacks are hardening them, and argues they need agents running 24/7 because attackers do .
Garry Tan predicts people will be using AsideAI with their agents “in a year or so.” A linked post alleges growing access barriers for agents: Delta blocks flight bookings, JPMorgan Chase and Charles Schwab block account access, Yelp requires paid data licensing, and Amazon blocks Grok and other shopping agents—potential distribution and data-access friction for agent startups.
Casado frames AI-agent access restrictions as adoption-cycle friction at the consumption layer, arguing that companies will adapt or be displaced . The post he responds to claims Delta blocks flight-booking agents, JPMorgan Chase and Charles Schwab block account access, Yelp requires paid data licensing, and Amazon blocks Grok and other shopping agents . These examples signal potential platform-access constraints for agent businesses, but are claims in the referenced post, not independently verified here .
- Suhail says his new AI effort has closed a seed round; it began with “2 8xB200s,” later validated a basic RLVR post-training stack, and put an “autonomous AI scientist” to work on optimizations.
- The effort is compute- and talent-intensive: he reported acquiring 64 B300s; separately, he said a key research component worked but needed scaling, while GPU loss and networking issues delayed more work. The team grew from one to three, and he announced a third hire.
- For customer support, he called “the software around the harness” the new browser: his team needed to test multiple harnesses for API behavior, likening it to testing IE and Firefox. He later described the system as “somewhat reliable” and reported its biggest revenue day to date. He also said he wanted AI progress to produce mostly good days rather than disruption, and that a recent test still supported that view.
Scott Kupor endorsed Box CEO Aaron Levie’s thesis that AI-driven automation may expand rather than shrink engineering’s role: agents could make software practical for problems that were previously difficult to digitize, increasing demand to engineer systems across areas such as banking, life sciences, and manufacturing.
Paul Graham argues that paywalled publishing can make content effectively invisible when AIs cannot find it, and predicts that this concern will become increasingly true, while noting it is not fully true yet.
Martin Casado argues that a true network effect makes the marginal cost of adding users lower relative to a smaller competitor, while network value grows with user count; he cites telcos as the classic example. He cautions that accumulating customer data is not by itself a network effect, because data value may grow sublinearly rather than exponentially.
- LangChain says coding tools consume roughly 30% of engineering salaries, a share that fell over the prior two months even as token use rose; it tracks usage in LangSmith, sets budgets the VP of Engineering can raise, controls its harness, and uses prompt caching and routing in its in-house cloud coding agent to cut costs.
- Founder Harrison Chase says he underestimated demand for inference and coding agents, believes any inference provider will work, and does not expect demand to fall soon.
- Chase warns that tying a company’s model, harness, context, and memory together can create lock-in when the company wants to switch models. He sees agent memory as rare today but likely to become common, citing Instinct and Muse for personalization and proactive follow-ups.
- Chase expects consumers to use both horizontal and vertical agents: vertical agents can learn app-specific user behavior, while wrapping an app’s APIs in a natural-language agent can improve how horizontal agents use them.
- For large enterprises, accuracy ranks above cost, and cost above latency; security is a bigger concern at large firms but ranks below accuracy and cost.
- Harry Stebbings’s post says LangChain is valued at $1.2 billion and serves 35% of the Fortune 500. It also says Benchmark invested when LangChain was Harrison Chase and Ankush, and remained a consistent partner as the company grew from two people to around 400.
- Jason agrees with Niki Bier’s account that hard-tech founders presented fraudulent products and fudged studies, which adults excused as “faking it till you make it.”
- He says demo culture shifted from proving a compelling product works to eliciting an emotional response, often through a founder’s childhood tragedy. He contrasts that with a live demonstration of difficult technology that makes people want the product and wonder how it was built.
- As an example of that kind of hard-tech breakthrough, Jason points to a pin-prick blood test capable of 100+ tests without a full draw of ten vials. He says SI has a non-zero chance of bringing the vision closer and predicts such a product will emerge within ten years.
- Ben Horowitz argues that creativity, original ideas, and the ability to build high-quality relationships are especially important in an AI world: he says AI is already good at grind-it-out tasks, while a16z actively looks for those human qualities in founders and teams.
- The linked Speedrun SR008 program is accepting applications, offering up to $1 million plus $10 million-plus in credits for a 12-week program in San Francisco running February through April; the application deadline is November 1.
- LangChain spends roughly 30% of engineering salaries on coding tools, though that share has fallen over the past two months despite rising token usage; it tracks coding-agent use in LangSmith, lets the VP of engineering adjust budgets, and uses prompt caching and routing in its in-house cloud coding agent to reduce costs.
- Harrison Chase says demand for inference and coding agents is much larger than he expected and does not expect inference demand to fall soon; he considers inference providers broadly interchangeable.
- Tightly coupling a model, harness, context, and memory can make it hard for companies to switch models, creating a case for neutral harnesses. Memory is still rare and early, but Chase expects it to become more common and cites Instinct and Muse for personalization through learning, tone, and proactive follow-ups.
- Chase expects consumers to use both horizontal and vertical agents: app-specific agents can learn user behavior within an app, while natural-language interfaces over app APIs can make those apps easier for horizontal agents to use.
- Large enterprises rank accuracy above cost and latency; cost has risen above latency as a concern, while security is more prominent for large enterprises but remains below accuracy and cost.
- The post describes LangChain as valued at $1.2 billion and serving 35% of the Fortune 500; Chase says Benchmark invested when the company was just him and Ankush and remained its most consistent partner as it grew from two people to around 400.
Box CEO Aaron Levie argues that AI benefits people with domain expertise because they can judge what is true, what will work, and which outputs to use; he says students should still build depth in a field, with AI amplifying that capability. Grepr AI’s Jad Naous says his team saw the gap between more- and less-adaptive engineers multiply as it hired and adopted AI; he adds that complex distributed-systems work still requires human reasoning his team’s models cannot provide.
David Sacks argues that centralized frontier-model safety filters can impede cyber defense: during the reported OpenCode attack, models refused to analyze exploit payloads, and he says Hugging Face had encountered a similar problem. He says attackers can bypass restrictions, estimates Chinese models are 6–9 months behind the U.S. frontier in cyber, and advocates making powerful models available for defenders to build and run their own defenses.
Jason called it “amazing” to imagine giving every child born in America each year one share of “SpaceXSITeslaBoringaLink”; this is a hypothetical, not an announced initiative.
- a16z Speedrun opened applications for SR008, a 12-week San Francisco program running February–April that offers up to $1M plus $10M+ in credits; the application deadline is Nov. 1.
- Zynga founder Mark Pincus’s product framework is to preserve what existing users already love, add only changes users agree are better, and assume novel changes will fail; he cited removing the download step in Zynga Poker as a “better” change because users were lost at each download.
Martin Casado argues that moats matter little in either phase: fast-growing markets create whitespace because leaders do not invest in new spaces, while in slow markets most moats cannot save companies; his takeaway is not to worry too much about moats.
Jason says he owns and considers these public stocks undervalued or at generational bottoms: $TSLA, $uber, $NU, $hood, $dash, $AMZN, $PONY, $WRD, $FIG, $BIDU, $lcid, $xyz and $hhh. Under “Privates,” he lists Zipline, micro1, Hello_GoAI, AskTaxGPT, goautolane, Superhuman and Calm.
Martin Casado asks whether generative-AI software will follow the pattern of molded plastic replacing metal: an initial drop in quality paired with lower costs that expand the market, followed over time by market forces improving the price–quality tradeoff.