Summary

Deedy Das’s marquee portfolio bet is Goodfire, a mechanistic-interpretability lab he calls “brain surgery for AI models.” Founded by researchers from the early interpretability teams at Anthropic, DeepMind, and OpenAI, its premise is that all current explainability is merely empirical — GPT-4o’s sycophancy wasn’t really caught in evals — and Das teases a still-secret result: “they have done a big part of one pillar of the five things they’re looking to solve.”
His most tradeable macro frame: demographic decline means AI for unhireable jobs is a need-now market, not a bet on future capabilities. Kids want to be YouTubers and coders — “no kid is dreaming about being an accountant” — so insurance brokerages, trucking, and logistics struggle to hire, while the Valley mostly builds “AI for finance and AI for legal” because that’s where Ivy League kids went. “You’re not betting on some future where AI can solve math or coding… You need this now.”
“Almost no iconic company in the history of venture capital has come from anyone’s thesis area” — Das says VCs pitching theses are “semi-lying.” Invoking Khosla’s OpenAI investment, he says it wasn’t a thesis that LLM chatbots would be a thing; nobody chose Facebook over Hi5 and Orkut by framework. His real filter: founders who’d wake up in five years saying “I would be doing nothing else except this,” plus the ability to recruit with near-cult-leader conviction — the biggest bottleneck he sees.
Probe deeply and “at least 90% of companies are actually just changing a prompt on an LLM.” He watched the same founders swap “data flywheel feeds back into fine-tuning” for “it’s RL” a year later, then go silent when asked what RL means; he also says at least 70% of AI pitches reduce to “I just wanna build models. I’m not sure why it’s useful.” For genuinely deep teams, Carta’s 30% premium for AI-enabled or AI-named Series A companies can be fair — “I would 100% pay that premium” — though Molly’s steelman is that the premium for actually good companies may be much higher.
Over-raising can bend the capital-velocity curve negative: illusion of success, the “$50 lunch problem,” and recruiting that can’t show upside. Das says he can’t justify a billion-dollar seed from his own fund; one explanation is an SPV where the lead puts in $1 million of a $100 million round and is “effectively taking none of the risk here… just getting all the marketing value.” Founder secondaries create another bad incentive: “I kinda got the bag.”
His infra map: pretraining is data-bound, so we’re in the RL era — Mercor and Turing filling the hole Scale AI left after its acquisition — with unresolved bottlenecks beyond it. RL is “kind of a shitty paradigm for learning” (reward only arrives at the end, a point Karpathy discusses), and the open problems include sample efficiency, agents/test-time compute, memory, determinism, and context windows; his north star is Anthropic’s “economic Turing test.” His honest hedge: “I do not know the future. I do not know how to answer things like AGI 2027.”
He endorses the layoffs-as-fitness thesis bluntly: “most engineers don’t do shit.” Glean’s Arvind told him he knew Google ICs who hadn’t written code in ten years; AI-attributed tech layoffs are mostly a post-zero-interest-rate efficiency correction, with transitional pain but historically new work emerging. His best inherited wisdom, from Arvind: at any given moment only one question matters — “do customers love this product? Is the answer yes or no?”

Deep dive

1. Goodfire: “brain surgery for AI models,” with a classified breakthrough

Deedy’s marquee bet is Goodfire — his gloss is “brain surgery for AI models,” technically mechanistic interpretability — built by researchers who founded or worked with the very early interpretability teams at Anthropic, DeepMind, and OpenAI. The thesis: “there is no future that we wanna live in where this is a black box,” and all current explainability techniques are empirical — they evaluate outputs rather than “fundamentally explaining why a model does what it does.”
His load-bearing example: GPT-4o’s sycophancy episode — “that’s not something they really caught in eval” — versus peeking into the model’s brain to see “what it was actually thinking, why it was actually saying the things it said.”
On backing labs: they’re not on the “revenue train”; the anxiety-inducing question is which fundamental discoveries to chase and “when is the right time to productionize them.” The tease, delivered as “can’t say”: a “super niche scientific breakthrough” that has completed “a big part of one pillar of the five things that they’re looking to solve.”

2. “Early innings” means demographics, not hype — and fertility is the 50-year question

Deedy admits “early innings of AI” is the most cliché VC line, then defines it precisely: birth-rate trends mean “there are just a bunch of jobs that people can’t hire for anymore” — nobody wants blue-collar or accounting work — so AI must fill the gap. “You depend on it… You’re not betting on some future where AI can solve math or coding in a specific way. You need this now.”
The Valley’s blind spot, as he tells it: “Talk about finance and legal, ‘cause that’s where other Ivy League school kids go and work. Therefore, there’s an AI for finance, and there’s an AI for legal.” Meanwhile insurance brokerages, trucking, and logistics go unserved — “just because you know a couple of private equity guys and a couple of lawyers doesn’t mean that’s all everybody does.”
His longer-horizon curiosity is fertility, explicitly not “this pronatalist agenda.” Modern humanity has never seen systemic population decline. Citing some smart people who argue that consumption and economic growth depend partly on population growth, he asks, “where is spending going to come from?” and calls it an interesting 10-to-50-year question.

3. The X playbook: a helpfulness bar, thick skin, and one five-minute phone call

Two failure modes he sees: people want Twitter because “they like other people hearing the sound of their own voice,” and the “overachiever archetype” — studied all their life, went to Stanford — “can’t take Twitter when it gets real. ‘Cause you can’t do Twitter without getting canceled a few times.”
His origin was writing and data exploration: “Hacking the Indian Education System,” a freshman-year scrape showing that India’s equivalent of an AP exam or SAT, taken by 1 million people annually, had statistical anomalies in its grading. The formative moment came from college-admissions AMAs: a girl in India, admitted to Penn M&T with financial aid, whose father wouldn’t send her. Deedy, an uncomfortable 19-year-old, spoke to the dad for five minutes — “I’ll send her to Penn.” She graduated and works in private equity. Lesson: “the asymmetry in… true influence you can have on people’s lives by just sending a text is so high that it’s insane that more people don’t do it.”
The operating filter now, after big tech “threatened to fire me a couple of times for brand reasons”: every tweet must pass “is this helpful for somebody?” — about 90% do. Data content stands out because “it is literally fact… you can’t argue against the fact.” Tooling: a custom Claude Code skill encoding his exact chart aesthetics, plus Apple’s Freeform for composites.

4. The $100M Anthology Fund and Menlo’s low-volume, operator-heavy model

The Anthology Fund is Menlo’s $100M vehicle with Anthropic, set up “a decade ago in the AI world, but… at the beginning of last year” when Anthropic was still a no-name company. It was deliberately not a corporate venture fund. Three buckets: fantastic early-stage AI teams (checks from $100K up to leading rounds), companies of strategic importance to Anthropic (Turing, Mercor), and iconic companies building on Claude at any stage.
Menlo’s main fund is deliberately low volume — not even five deals per partner per year, with Deedy arguing that even three or four can become too many — because at that pace, two years in “you’re gonna have six companies that you’re heavily involved with, and you’re not really gonna have time to do work for any of them.” The unglamorous value-add: when a company stalls, “who’s gonna help you do an M&A motion? Most founders haven’t done that before… That’s when we come in.”
The differentiation is tenured operators: Tim Tully (Splunk CTO), Joff (Atlassian chief product officer), and Matt Kroening (one of the few cybersecurity unicorns sold above $1B in 10-15 years). “We’re not just investors. We could build this company with you.”

5. “All of them are semi-lying” — the anti-thesis thesis and the founder sniff test

The hot take, delivered with named receipts: “almost no iconic company in the history of venture capital has come from anyone’s thesis area.” Invoking Khosla’s OpenAI investment, Deedy says nobody was investing on a thesis that “LLM chatbots are gonna be a thing”; nobody picked Facebook over Hi5, Orkut, or Google Plus by framework. “The investment is, these guys are really smart, and they’re working on something that could be pretty valuable. That’s your thesis.”
His number-one founder screen: “are you just doing this because you think being a founder is cool? ‘Cause everyone thinks being a founder is cool and high status. I don’t care about that kind of founder.” What he needs to believe: “I want you to wake up five years later and be like, ‘I would be doing nothing else except this.’” He says status-motivated founders aren’t necessarily bad and often do well, but he wants to avoid the founder who blows the raise and shrugs.
Conviction requires time — “you just can’t trust the nature of a founder when they’re pitching to you right before a raise… They’re selling to you.” He builds relationships pre-raise, half the time not about work. And even mission-true founders fail his second gate, hiring: “you almost kind of need to be a little bit of a cult leader as a founder.”

6. Boring verticals, frontier research, and the lost art of Granola — plus the Cluely footnote

Three areas. First, the information-asymmetry verticals — ideally founders who grew up inside an industry (“maybe your mom, maybe your dad worked in it”) rather than backing in via top-down research. Second, research risk: “research is not happening in academic institutions anymore. They just don’t have the money to fund frontier research” — so who underwrites high technical risk with economic unlock? Third, beautiful products: “no one is waking up every morning like, ‘I need to fund a meeting note-taker company’” — yet Granola works because “it just works… That’s a lost art.”
The PLG law attached to that taste thesis: “if your enterprise product can be PLG, then it must be, otherwise a PLG company will absolutely eat your lunch and destroy you.” Enterprise sales cycles run slower than the technology — you make irreversible product decisions (“classic RAG solutions have now become agentic solutions”) — and then the viral product walks into your sales call: “the buyer is like, ‘Yo, I’ve heard of this one. We should buy that one.’”
The Cluely story fits here: when it was still Interview Coder, Deedy broke his own Twitter rules — “Cheating is really bad. Absolutely don’t do it… This is the tool in case you ever use it by mistake” — a post with several million views, the first time anything related to Cluely or Roy went super viral. He defends Roy as “a very thoughtful guy.” His real grievance: “strivers” who gamed tech interviews into director seats — “you guys know nothing about building product. How are you a director at X company?” Within a month, he was the one asking Roy how he does it.

7. The 30% AI premium is justified — for the minority who aren’t just re-prompting

From 10-20 pitches a week, he’s watched buzzword fashion cycle: “our differentiator is that we get data, and then data flywheel feeds back into fine-tuning” became, from the same founders a year later, “that same thing, but instead of fine-tuning, it’s RL” — followed by silence when he asks what they actually know about RL.
His edge as a technical investor: “we’re not embarrassed to ask the dumb questions. A lot of VCs are, ‘cause they don’t code.” Probe deep and “for at least 90% of companies, they’re actually just changing a prompt on an LLM,” pitching futures they hope exist. He also says at least 70% of AI companies fall off because “the TLDR of their pitch is, ‘I just wanna build models. I’m not sure why it’s useful.’”
Given how constrained the real talent pool is, Carta’s 30% Series A premium for AI-enabled or AI-named companies is defensible for genuinely deep teams, especially at seed and A within his broader seed-through-B focus: “we’re betting on an outcome that’s a 10X, 100X… I would 100% pay that premium.” Molly’s steelman — worth keeping: the premium for actually good companies may be “much, much higher,” because once competition arrives “pricing goes out the door.”

8. Over-raising: the curve inverts, and some billion-dollar seeds are marketing

His capital-velocity curve: raise $15M and get ~80% of the achievable acceleration, $100M maybe 90% — but often the curve actually turns down. Three mechanisms: the illusion of success (“We’re already a billion-dollar company. Can we even fail?”), the “$50 lunch problem” of fancy spending killing hunger, and recruiting — “how do other people see the upside when you’re already raising at whatever price without any product?”
He says he can’t justify a billion-dollar seed from his own fund; one explanation is an SPV structure: put $1M into a $100M round, raise $99M from others, “I’ll put my name on it, but I’m effectively taking none of the risk here. I’m just getting all the marketing value of being the big, bold VC.”
The uglier incentive is founder secondaries. “The founder basically exits and goes like, ‘Well, who cares if I over-raise?… I kinda got the bag.’” Possible exceptions he gives include making a market splash, competing on paper equity value against Anthropic paying “$4 million a year, or whatever amount they want to pay,” and consumer companies using the money to buy users. Meanwhile venture ignores companies with hundreds of millions of ARR that many people in core tech Twitter simply don’t know because they “grew slowly over 10 years.”

9. The infra arc: internet data hits a ceiling, RL ascendant, several bottlenecks beyond

His compressed history: GPT-3 showed internet-scale pretraining generalizes; RLHF meant “we basically sort of solved the Turing test overnight”; scaling laws drove the big-model phase until the available data became the constraint — “order of magnitude-wise, you’re not 10X’ing the data.” Hence the RL era: buying domain data (code, Excel/finance) with Mercor and Turing filling the gap “as Scale AI left a big hole after the acquisition.”
What’s unresolved: RL’s scaling laws “are not intuitive because you’re depending on this manual process of getting data” — and per Karpathy, “RL’s kind of a shitty paradigm to learn… I only know at the end that I played a shitty game.” Beyond that: better learning from fewer samples, agents chaining tools under test-time compute, memory, reliability (“they have no core structural understanding of things”), and context windows.
The framing he likes as an end goal is Anthropic’s economic Turing test: “can I pay an AI X amount of money to do a task that I could pay a human to do as well, and I would not know the difference? And can I drive that value of X higher and higher?”

10. Investing without knowing the future — and “most engineers don’t do shit”

His honest non-answer on capability forecasts: some believe the hours-of-work-per-model curve keeps rising toward automating most human work; others say current techniques produce only “pseudo-novel” discoveries — “a twig on one side” of the knowledge tree, never new branches. “My TLDR to that is I do not know the future. I do not know how to answer things like AGI 2027.” So he reasons from present facts — hence OpenRouter: “one API could give you access to all of them… It’s not rocket science-y at all… that’s obviously something we need.” Asked to predict the best model at year-end, he defers to the market: “It’s Google.”
On Molly’s relay of Coatue’s Michael Barton’s take (layoffs as a health indicator, “getting fit”), Deedy goes further: “most engineers don’t do shit. We know this… we know how many trips to Hawaii a year they go on.” His Glean-kitchen anecdote: complaining about people at Google who hadn’t coded in a year, and Arvind smiling — “A year? I knew people who haven’t written code for 10 years, and they’re ICs.” Elon “sort of proved this with Twitter.”
His net on AI and jobs: tech layoffs are more “a laziness, zero interest rate phenomenon” correction than AI displacement; across history “some jobs will go away. People who have the jobs will be very upset… but in the long run, humanity finds a way” to find new things to do — with real transitional pain, like the Industrial Revolution factory worker “who was 35 with three kids… I hate robots.”

11. Three lessons from Arvind Satyanarayan: hard work is a gift, keep the cool company, one question at a time

The person he admires most is Arvind Satyanarayan, whom he worked with at Glean — a multibillion-dollar paper net worth, works harder than anyone at that level, “drives shitty cars,” hates podcasts and events, and told him: “Don’t complain about hard work. Hard work is a gift… you don’t know how lucky you have it.”
Facing a serious acquisition offer that would have let Arvind be out with $1B liquid — while asking Deedy to drive him up 280 to avoid paying for an Uber — his answer: “I run such a cool company right now. If I get acquired, I will no longer run that cool company. Why would I want that?… I don’t wanna be some exec schmuck at some company that bought me.”
The clarity lesson, when early-Glean Deedy peppered him with questions about defensibility and sales ramp: “You have a lot of questions. They’re pretty good questions, but they’re not the right question. At any given point, you have one question you’re trying to answer… The one question we have right now is: do customers love this product? Is the answer yes or no?… None of that other shit matters.”

Bill Chen's follow up on REITs and $ALX She Raised $64M to Build an AI Math Prodigy | Carina Hong, CEO of Axiom