prev ceo @tensorstax (acq by @snowflake)
sf
Joined May 2023
- Tweets993
- Following620
- Followers651
- Likes8.1K
Pinned Tweet
Very delayed news, but back in November @tensorstax was acquired by @Snowflake.
Since then our team has been hard at work building Cortex Code, the current SOTA data agent.
I'll be sharing more of our work/research going forward around harness engineering, cloud sandboxes, eval environments, RL and much more.
aria retweeted
Nothing derails a young man more than a near bag experience
Yet in pursuit of a bag, a near bag experience is almost a prereq
Incredibly interesting read
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/
aria retweeted
A year ago, many dunked on Meta for spending billions to hire top AI talent and acquire Alexandr Wang (I was skeptical too, tbh).
But the success of the Muse app and models proves one thing:
Hire top talent, assemble a cracked team, give them an ambitious goal and enough resources, and they can work wonders.
The biggest consumer of personal agents will be enterprises and most likely the interaction surface will be slack and teams
aria retweeted
If you believe the risk is coming from 1-3 people in a garage with no money and no compute, you just don't understand this technology and haven't learned anything this summer.
Risk comes from the asymmetry of capabilities created by secret labs training frontier agents and running them with massive amount of compute. Open-source is exactly the solution to this asymmetry and empowers hospitals (and any smaller orgs) to defend themselves!
We're used to thinking of open-source models as an unadulterated good. But in the case of AI, they can actually pose additional dangers, as @ReidHoffman and I got into at #CGI2026. I appreciated this nuanced discussion.
Some work for the sake of working, others work for the sake of getting the job done.
Never stop moving until the job is done.
aria retweeted
Another example from the Snowflake earnings call of customers building workflows and applications directly on top of their data
Customer essentially rebuilt Gainsight in 5 hours
What if Snowflake isn’t a victim of the SaaSpocalypse but rather the assailant…
aria retweeted
This is what we have been working on for the last 6 months or so at the AI Snowflake Research:
Zero Redundancy Rollouts (ZoRRo):
snowflake.com/en/blog/engine…
If you do RL and you want it to be much faster make sure to have a look.
aria retweeted
If this is true, using the best public estimates we have of LLM resource use, solving this Erdos problem took 0.6–6.3 kWh of electricity and about 3–31 liters of water.
So that is less than three almonds worth of water and the electricity equivalent of 2-20 miles of EV driving.
just quick napkin math on how long this took (unless i missed where they said):
the published CoT summary is 111,145 tokens long. it's really hard to say how much they summarized, assume 3x-20x reduction in tokens?
and i'm assuming this is gpt-5.6 pro, so taking Artifical Analysis' benchmark of 51ms tok/sec at 100k input for gpt 5.5. underestimate prob hard to say
this seems a bit low so going to multiply all of this by 2x
then this probably took anywhere between 5 hours to 32 hours. so like $120 - $1000 in gpt 5.5 pro tokens
whole point is not that long for a result of this magnitude!
More details here:
snowflake.com/en/blog/tensor…
aria retweeted
I trained models across MacBooks using Apple's AirDrop protocol.
grove is a distributed training library for Apple Silicon. Devices discover each other over AWDL, a direct radio link. If there's a shared WiFi network it upgrades to that for speed, otherwise everything goes over the direct link. No router, no cloud, no setup.
grove start <script> -n 4
grove join
Taste will continue to be a very long lasting moat.
Today we're launching Glaze 💠
Create any desktop app in minutes by chatting with AI.
Beautiful, powerful, and truly personal.
Learn more on glazeapp.com
Follow @glazeapp for updates.
aria retweeted
imagine running @openclaw on one of these
Starcloud (@Starcloud_Inc1) recently made history by launching a satellite with an NVIDIA H100 into orbit — the first time a GPU that powerful has ever operated in space. It's the first step toward building AI data centers in orbit, powered by continuous sunlight and cooled by radiating heat into deep space.
Their approach could one day rival the world's biggest data centers while using less energy, zero fresh water, and far lower emissions.
YC's @aaron_epstein visited Starcloud's HQ, where co-founders @philipjohnston, @EzraFeilden, and @Adi_Oltean explained how they built a working prototype in just 15 months — and why big tech is racing to space for AI compute.
