frontier research & venture @frontiersynd certified futurist (IFTF) | agentic infra @constructive_io SF lead @vetsintech | USMC | prev @accenture R&D

San Francisco, CA
Joined July 2013
the GPU market is getting tight, fast Ornn's B200 rental index is up nearly 13% in a single week. IREN says pricing on new 3-year contracts has more than doubled since last november if you're training your own models or scaling inference, the window for cheap, flexible compute is closing. the teams locking in reserved capacity and colo now will have a real edge if you need GPUs, colo, or power, I work with @frontiersynd infrastructure partners who still have capacity DMs open
1
2
75
What if you trained a model to read a cell like a sentence of genes? Akram Baharlouei @yalda2009 of Altos Labs tested that at FRC #16 at @Stanford. In her team's benchmarks, those BERT-style foundation models didn't do well against simpler, even linear, models. Small flow-matching models that skip compressing the data did better. Tonight FRC #28 is back at Stanford. Ganesh Pusapati on the broken machine behind most colorectal cancers. Angela Oliveira Pisco @drAOPisco on what a virtual cell can actually predict. Full talk 👇
2
1
1
82
better reasoning doesn't automatically mean better actions. @joseph_lee_25 showed both sides at FRC #27: fewer collisions on a driving model, almost no change on a manipulation robot turns out, correcting a robot's reasoning only pays off when its actions actually depend on it
If you fix a robot's reasoning, does it act differently? @joseph_lee_25 from @Stanford tested it at FRC #27 at @MissionRobotBay. Steering a driving model's chain of thought cut collisions 30.4%. On a manipulation robot, better reasoning barely changed actions. Full talk 👇
2
111
same pattern everywhere: the applications are disappearing. the application LAYER isn't. it's just becoming pure intent: who can do what, with which data
I have stopped buying all SaaS since Q3 2026. The only software I bought include: - GCP/AWS for database, logging, deployment; email services etc. - Tokens, tokens, tokens - Database like customer enrichment data source - Connectors app that help me connect to other apps(e.g. Pipedream etc.) - I use Codex/ClaudeCode/Hermes ... but pay for tokens Anything with complex UIs have been unsubscribed: - Any design/editing/video editing software (Capcut, Canva, Photoshop, Adobe everything) - Any docs/table/sheet/ppt software (Word, Powerpoint, Excel etc.) - Any UI-facing dashboard/analytics/admin/CRM software - Anything I have to watch a tutorial video during onboarding The only software I tolerated clicking and typing UI: - Collaboration with human: Lark/Slack etc. - Social apps: Twitter/Instagram/Linkedin/Wechat
3
90
Last Wednesday, Frontier Research Club #27 filled @MissionRobotBay for Robotics & Embodied AI: Feel or Think? Joseph Lee from @Stanford showed that steering a driving model's chain of thought cut collisions 30.4%. On a manipulation robot, the same fix made the reasoning better and barely changed what the robot did. Shashank Tripathi from @MPI_IS made the case for feel: learn contact from the way people already touch the world, reconstructed in 3D from ordinary images, then transfer it to robot hands. Great to see everyone in the room: @NRaghav_4 @ja_rothschild @shenyshawn22 @EdmondIsARobot @m_goes_distance @gonghuachao @broodsugar @pushkalkatara @pranaypasula @KateYXu @keertimanney @aditya_potnis @parassavnani @khurram_pirov @hynwprk @JasonMeaux @OrriBogdan @SabashanRagavan @AlyssaTsai @deseventral @chrisnolet @paramendra @guru3s @OFR92 @stsqit @Sapanostic @JolinJiao @GarrettCAllen @BenHolfeld @Dylan_Bodkin @dalehitt @shayadeaton Both full talks are up now 👇
1
1
4
209
Kristopher Floyd retweeted
If you don’t move fast, you fall behind. @GuralNuriyev highlighted the need for AI researchers to rapidly iterate and adapt. Grateful to @Kristopherfloyd for cohosting with us.
the gap between a lean team and a frontier lab is closing fast recently, chapter 9 resident @GuralNuriyev hosted the frontier research club for a night on recursive self-improvement together with @marcelluzs & @shoobiedoo313 gural is giving lean AI teams an army of research agents running parallel experiments, so they can ship frontier work in-house, continuously, without the overhead of a big lab
1
1
3
8
632
Last week I helped host the National Security + Space Innovation Summit with @VetsInTech and @BayAreaCouncil. Two days aboard the historic Klamath at Pier 9, during Fleet Week and SF Tech Week. The highlight: Artemis II pilot @AstroVicGlover rode @navierboat's Quanta-D foiling vessel on the Bay with the @BlueAngels overhead. He was the first test pilot from outside Navier's core team. Huge thanks to my friend @sampritibh for making it happen. Friday night we crossed to the USS Hornet for Victor's unofficial retirement party. I met him and his Artemis II crewmates @astro_reid and @Astro_Christina. Something I'll never forget. Some news: I'm now the SF Bay Area Chapter Lead for VetsinTech. Full write-up: newsletter.thefrontiersyndic…
1
1
7
4,933
Ganesh's result is the kind of wet-lab surprise models need to catch up to: the protein the destruction complex is supposed to destroy also helps hold it together. Angela's asking the harder follow-up. Looking forward to this one at Stanford.
More than 80% of colorectal cancers break the machine that keeps WNT signaling in check. Next Wednesday, Frontier Research Club #28 is at @Stanford for Cell Biology x AI: From Cancer Mechanisms to Virtual Cells. Ganesh Pusapati of the @RohatgiLab co-led a @NatureGenet paper that found ~150 previously unreported mutations in that destruction complex, then showed that strengthening one interface restored its function and slowed cancer-cell growth in the lab. @drAOPisco of @ChanZuckerberg takes the other half: what can a virtual cell trained on atlases like Tabula Sapiens actually predict, and what evidence would make you trust it for the next experiment. Dinner, two short talks, most of the night for critique. Wed Oct 21, doors 5:30 PM.
1
69
🤜💥🤛
If you don’t move fast, you fall behind. @GuralNuriyev highlighted the need for AI researchers to rapidly iterate and adapt. Grateful to @Kristopherfloyd for cohosting with us.
2
175
Most insane demo of all time
Real life, baby
2
183
seems like very good news
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works already-clean.goodfire.com/
2
128
AI is the new front end. i'm already incredibly frustrated when I have to manipulate software myself directly, instead of through my agents
David Sacks (@DavidSacks) says Marc Benioff is letting Claude own Salesforce's front end, and risking disintermediation on purpose: "He is integrating Salesforce into Claude. He's willing to allow Claude to be the front end, to have the sort of the primary user relationship." "So in other words, Benioff here is willing to risk disintermediation in order to ride this AI wave because he knows it's what the customer ultimately wants." "You don't want to go into Salesforce or any SaaS application to write your agents. You want to create agents in the AI program that you're using. And you want those agents to be able to run across all of your SaaS platforms." "Why is this ultimately good for his business and not bad is because, like I said, I still think that you need systems of record. You need the AI agents to basically go to those systems, get the data from a canonical source of truth. You just don't want there to be any probabilistic variability around that." "It's not a rip and replace." "Benioff is saying that the AI is now the user interface. He's basically acknowledging that he's leaning into it." - David Sacks, All-In co-host, on the All-In Podcast episode published August 29, three days after Salesforce and Anthropic announced their Claudeforce partnership.
1
2
93
I’m getting really excited about databases yall
34
Frontier Research Club #27 is tonight: Feel or Think? Robotics & Embodied AI. Does a robot need richer senses or better reasoning before it acts? Joseph Lee (Stanford, Somil Bansal's lab) on TRUST: monitoring & steering robot reasoning Shashank Tripathi (@MPI_IS) on DECO & PICO: learning from human contact Dinner + paper discussion at Mission Robotics, SF. 5:30 to 8:30 PM. Limited capacity, request to join: luma.com/sxoqc4og
1
1
104
Kristopher Floyd retweeted
Our CEO @danlynch opened FRC #26 with Secure-by-Default Postgres for Agents. Agents are database users now. Authorization has to live in the database, not the app. Thanks @frontiersynd @Kristopherfloyd @pebble_bed @danielchalef @zep_ai @shulynnliu @shagarw21
1
2
7
381
agents already create most new databases. soon they'll be most of the users too
Your next database user won't be a human. It'll be an agent firing off hundreds of queries a second. @shulynnliu and @shagarw21 of @BerkeleySky made the case at FRC #26 at @pebble_bed: agents don't query, they speculate. Schemas, samples, partial plans, retries, all in parallel. On BIRD, more attempts meant better answers. Across 50 parallel attempts, most of the work was redundant. Agents already create ~20x more branches and ~50x more rollbacks than humans. So the database has to change. Probes instead of queries. Satisficers instead of optimizers. Branching storage built for agents. Thanks to @constructive_io for sponsoring.
2
1
6
235
Reasoning VLAs are everywhere right now. Joseph's result is the one I keep coming back to: fixing the reasoning made a driving policy safer and did almost nothing for manipulation. Tomorrow we argue about why.
Should a robot think before it acts, or just feel more? Tomorrow, Frontier Research Club #27 is back at @MissionRobotBay for our Robotics & Embodied AI track: Feel or Think? Joseph Lee from @Stanford presents TRUST, a value model that reads a robot policy's chain of thought mid-generation and steers it. On a driving VLA, steering the reasoning cut the collision rate by 30% on a hard benchmark subset. On a manipulation VLA, it barely moved behavior. So when does reasoning actually earn its cost in a robot? Talk 2 brings the touch and dexterity side. Dinner, two short talks, most of the night for critique. Wed, doors 5:30 PM, Mission SF.
3
99
Twitter under @paraga measured bots with quarterly human audits, suspended 500k+ spam accounts per day, merged its anti-spam and anti-toxicity teams, and publicly exposed state-backed bot networks.
Zelensky on the effect of Ukrainian strikes on Yandex data centers: "When our strikes reached Yandex data centers, there was less hate, less division, and fewer provocations on social media. Because there were fewer bots, it's pretty clear where those bots were coming from."
78
Jevons is coming for databases too cheaper tokens don't mean less infrastructure. they mean more apps, more agents, and more databases. Supabase is adding 4M+ databases a month, and agents create 70% of them
I can't emphasize enough how important this @a16z chart seems and how few have internalized it.
6
201