Stealth startup founder. Neuroscience PhD from Stanford: silent speech interfaces. GP of Asimov Ventures: 3D printing, robotics, deep tech.
Stanford, CA
Joined June 2009
- Tweets1.5K
- Following850
- Followers2.5K
- Likes465
Pinned Tweet
Our silent speech preprint is live! Using a cross-modal training technique enhanced by LLMs, we set a new state-of-the-art for silent speech (12.2% word error rate, open vocabulary) and brain-to-text (8.9% WER; Rank 1 on Brain-to-Text Benchmark '24) arxiv.org/abs/2403.05583
ALT Anatomical drawing of facial musculature and a steam-punk mechanically-integrated silent speech interface
Tyler Benster retweeted
A list of stats from the Anthropic S-1:
- NDR is ~900%
- Revenue per employee is over $6.6M, up from about $1.4M a year earlier
- Revenue went from $41M three years ago to $16.3B in the first six months of 2026
- Quarterly revenue went from $4.7B in Q1 to $11.6B in Q2
- Q2 2026 was the first profitable quarter: $974M in operating income ($1.7B operating loss the quarter before)
- Adjusted operating income was $1.6B in the first half of 2026, up from a $1.8B adjusted loss a year earlier
- Revenue grew 1,368% year over year in Q2, while compute and infrastructure costs grew 300%
- Compute went from 251% of revenue to 48% in five quarters
- The value of Anthropic’s own internal Claude usage grew 250x from January to September 2026
- Anthropic spends more on tokens (at list price) than on payroll
- Over 1,900 customers pay $1M+ a year, and about 250 pay $10M+
- Customers paying $100K+ went from 950 to over 11,900 in a year
- Over half of the Fortune 500 pays $100K+ a year
- Those Fortune 500 customers still spend less than 0.1% of their total expenses with Anthropic
- The top 5% of Claude Code users spend roughly 20x what non-users on Claude Enterprise do
- About 40 million smaller customers
- $72.4B of the $80.7B first-half net loss is an accounting charge on convertible notes (not operating losses)
- Nearly $800B in compute commitments over 10 years
$141.1B raised in total, with over $130B in cash and short-term investments
- Nearly half of revenue comes through Amazon, Google and Microsoft’s cloud marketplaces, all of which also compete with Claude
- A single share held by the seven co-founders carries 50.1% of the vote
- That share is redeemed for nothing if fewer than three founders remain
Krishna Rao is the CFO of Anthropic, and this is his first podcast appearance.
He joined the company two years ago when run-rate revenue was about $250M. Today it is $30B. He has helped raise ~$75B and is responsible for the procurement and allocation of compute.
I feel lucky we get to hear what it is like to sit inside a company this consequential at a moment this pivotal.
We discuss:
- The cone of uncertainty
- How he allocates compute across Trainium, TPUs, and GPUs
- What investors misunderstand about model companies
- Why the returns to frontier intelligence keep rising
- Platform vs application and where Anthropic builds its own products
- How Anthropic uses Claude internally
I have asked my closing question about the kindest thing more than 500 times. Krishna's answer is one I have never heard before.
Enjoy!
Timestamps:
0:00 Intro
2:38 The Compute Canvas
6:51 The "Cone of Uncertainty"
11:58 Why the Returns to Frontier Intelligence Are So High
16:45 Recursive Self-Improvement
20:20 Scaling Laws
23:30 Sourcing $100 Billion in Compute
28:05 Platform vs. Application Strategy
32:52 Pricing Dynamics
38:48 How Anthropic’s Finance Team Uses Claude
43:24 Raising Capital & Overcoming Investor Skepticism
52:32 Public Perception, Risks, and Government Regulation
57:25 Mythos Release
1:12:33 What Could Derail the AI Revolution?
1:13:47 Biotech and Healthcare
1:15:31 The Kindest Thing
Tyler Benster retweeted
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence.
block.xyz/dessa
Something I struggle with.. disregarding the true plight of the human labor displaced for a moment, isn't the field of Mathematics where papers are published with high velocity and with formal verification strictly superior to the previous paradigm?
nytimes.com/2026/10/16/us/ma…
Tyler Benster retweeted
Our house in Toronto had this fancy glassed wine cellar, but we are not big wine drinkers.
Anyways, it had good climate control so here is my rack cellar.
Congrats to @neurable on launching their new headphones! Can’t wait to try them. They look beautiful and I’m impressed by the persistence in making EEG work outside the lab for consumer. neurable.com/blog-posts/neur…
Tyler Benster retweeted
Andrew Huberman on why Meta, OpenAI, and Anthropic are all going to become biotech companies.
"Dario came up through neuroscience. Sam, I've spoken to, is extremely smart about the neuroscience, as is Zuck.
The technologists who used to just engineer stuff, they're becoming neuroscientists...and I think most people don't realize that.
They will all tell you, 'Okay, here's the goal. The goal is to be able to read and write from your brain non-invasively.'
Even just for things like drive and motivation, a non-invasive device where you could literally crank up your level of motivation by stimulating the nucleus accumbens and related pathways while engaging in a particular learning task or in hard work, and then turning it off so that you can capture more energy.
That's coming, and those companies are all about that. That's what the real arms race of AI is about."
My conversation with Andrew Huberman (@hubermanlab).
I didn't expect us to spend so much time on the frontier of neurotechnology, and how we'll soon be able to precisely measure and change our brain states, both awake and asleep. It was fascinating.
Also discuss:
- Building and keeping energy
- How to downshift and fall asleep quickly
- Neuralink and BCI
- The state of health media
- Following your curiosity
This was a fun one. Enjoy!
Timestamps:
0:00 Intro
1:00 Building Energy for High Performance
7:00 How to Fall Asleep
18:10 What’s Gone Wrong With Health Optimization
24:10 AI Companies Could Become Biotech Companies
30:45 Implantable Devices
33:45 Memory
42:50 Controlling Your Brain
59:00 Genetic Engineering
1:11:05 The Key to Meaningful Success
Congrats to @KarlDeisseroth, Peter Hegemann, and Georg Nagel for winning the Nobel Prize. I was in Karl’s lab for five years during my PhD, surrounded by the smartest and most driven people I’ve ever met. I learned so much developing techniques for whole-brain optogenetics
The big surprise here is how poorly constructed the Nemotron datasets are. Clearly benchmaxxed models
When ingesting this much data, we want to make sure we don’t accidentally train on test (we’ve unfortunately done it before…)!
We filter the entire corpus for any 13-gram overlap with our evals (including our development set evals, since they guide our research decisions!).
ALT Matrix showing training-data overlap with benchmarks, displaying benchmarks 1–30 of 53. Rows are data sources and columns are benchmarks. Larger red squares indicate more distinct overlapping documents. Bars on the right show documents removed from each source; Nemotron SFT General has the largest total, about 38,940. A document can overlap multiple benchmarks, so square counts should not be added together.
Tyler Benster retweeted
Sliced my thumb open today. Tendon repair surgery next week. It's forcing me to go all in on Wispr Flow and voice and I love it
_My family does not love it_ 😅
I'm very loud and obnoxious 😭
We need a solution for thinking into computers. This is so much better than typing and yet it totally does not work in a world of humans. If we were all yelling at our computers next to each other, the world would be significantly worse and fail
Who is working on this problem? I need silent speech
Tyler Benster retweeted
Thrilled to announce @positron_ai's $875M Series C at a $5B valuation!
This round was co-led by @NEA, @Atreidesmgmt (@GavinSBaker), @valor (@AntonioGracias), @AndraCapital, @SemiAnalysis_ Capital (@dylan522p), and Jim Clark (Founder of SGI and Netscape). An incredible group to have backing what we’re building.
Also very happy to have strategic backing from VentureTech Alliance, Hudson River Trading, Cisco Investments, and Naver Ventures, alongside continued support from @Arm , @jumptrading , @dfjgrowth, Qatar Investment Authority, and many more.
What I’m most excited about is what this lets us do next. We’re deploying more than 50 racks of Atlas at @OracleCloud, and this financing funds our next-generation Asimov chip’s tapeout, the production ramp for Titan, and the manufacturing capacity to meet the demand ahead of us.
The goal remains simple: make running AI dramatically cheaper and more energy efficient.
A huge thank you to the entire Positron team, especially @mitesh711 and @kmett, our customers, partners, and investors. Building hardware takes an enormous amount of work, and I couldn’t be more proud of this team. We have a lot ahead of us, and I can’t wait to share more.
Tyler Benster retweeted
Hedge funds will eventually provide inference to model routers for free, or even pay to do so, in order to get a first look at prompt contents. Payment-for-order-flow in the age of compute.
> The woman, Fiona, is wearing a NEWSOM/CLAUDE 2028 t-shirt; the man, Istvan, a matching VANCE/GROK 2028 one.
astralcodexten.com/p/the-foo…
Tyler Benster retweeted
When is @googlemaps + @Waymo going to roll out a parking spot locator? With the number of Waymos in SF, it feels like the coverage would be high enough to have a useful feature.
props for open science, but this isn't open-vocab, it's top10 retrieval from a closed set of 512 candidates
useful missing control: how well does Llama rank the true next word among those candidates using only the preceding ground-truth text (no EEG at all)?
Today at @nubrain_io we released a preprint of our research: decoding silent reading from dry, non-invasive EEG.
~240k word presentations · ~49h EEG · open vocabulary · scales with data
Paper: arxiv.org/abs/2608.20186
Rather than teach students to be self sufficient when cut off from AI, perhaps tests should allow use of AI but get a lot harder.
economist.com/graphic-detail…
Tyler Benster retweeted
If an alien visited earth and asked you: “Show me one chart that sits upstream of everything your species has accomplished or will ever accomplish.”
What would you show them?
I would show them the chart below.
From my perspective, as long as we keep this chart going, we will figure it all out…
Tyler Benster retweeted
Introducing terminal-code: VS Code inside the terminal
- VS Code compatible CLI
- works over ssh
- syncs with your terminal theme
terminal-code.com
Tyler Benster retweeted
Alibaba's open-weight AI models accumulated 3 billion global downloads in the past 6 months — more than Meta Llama, more than Google Gemma, more than every Chinese domestic competitor combined. The most-downloaded AI on Earth right now was built in China and given away for free. Think about what that means for the next 5 years.