builder of many things // I lead tech @ resin8 // cohost of @ai_rebels
Joined June 2021
- Tweets14.2K
- Following1.4K
- Followers4.6K
- Likes21K
∿spencer. retweeted
We've developed a new inference engine, called 'e'.
It's the closest thing to alien technology which we have, and pushes 10,500 tok/s aggregate on a single DGX Station with the new @NVIDIAAI Nemotron 3.5 Super VL model, released a few minutes ago.
Actual Computer turns your NVIDIA DGX Station into a department-scale AI factory.
Read the full case study:
actual.inc/company/blog/case…
∿spencer. retweeted
Someone could make a *lot* of money creating a curated influencer network of people who like to post about AI tinkering, put them all in touch with a roster of interested companies, and exchange resources like compute/credits in exchange for marketing they're already doing.
∿spencer. retweeted
i applied to cloudflare and they asked me if I would still work there if the company went to 0.
they also asked me if i would donate “glowing bile” to the lava lamps.
i am not exactly sure what the second question meant but I said yes to both.
∿spencer. retweeted
Actual Computer is pleased to announce its first source available-software, toks.
Actual's toks is the best tokenizer on earth and the first tokenizer designed for the quadrillion-token era.
actual.inc/company/blog/intr…
∿spencer. retweeted
if you're the one pitching, you send the invite.
i scheduled meetings with hundreds of prospective LPs for pitches. I never once asked them to send the invite.
from a founder:
∿spencer. retweeted
oh ABSOLUTELY i do not disagree with this AT ALL
true knowledge will dominate the world even more LMFAO
the models are defaulting to simply 'good code' in their training data and RL
they do this by fact checking their work, tests, and actually doing QA in the app. their cheat rate has gone down SIGNIFICANTLY in the last year
as long as the agents have a feedback loop to prove what they're doing the NEWEST models will make good system designs as a base trait, especially if they're encouraged too
∿spencer. retweeted
you can just run a Jev-like model with Actual on your own devices btw at FP16
I put one on models.actual.inc
∿spencer. retweeted
This is THE best possible time period in history to disconnect from the norm (like corporate jobs) and find success doing what you love!
Legit focus on your passion and learn how a team of ai agents can help you make money doing what you love
it's possible. I promise!
∿spencer. retweeted
this is vintage 2009 era internet content and i am requesting more of it on the timeline, thank you
∿spencer. retweeted
if you’re into harness engineering please get into multiplayer game programming patterns so we can stop the slop and get better inference utilization by deploying a fleet of little guys who can all talk and see each other and do whats best for the team at any time ty
∿spencer. retweeted
I think most inference ends up being free for consumers
If inference capacity goes where we all think its going
The human feedback for the inference will be much rarer
Eventually the interaction ends up valuable enough to cover the inference and then some
∿spencer. retweeted
the reason nobody does evals is because most tasks aren’t literally deterministically verifiable but everyone including the models somehow got the idea that “LLM judges don’t work” and so people feel like it’s impossible
jev fixes this, culturally speaking
∿spencer. retweeted
Assume misalignment. Treat agents like extraordinarily useful malware.
∿spencer. retweeted
This has been my argument for a while. AI will benefit defenders long term.
If the best models in the world today can't find any more issues today, you're probably in pretty good shape against AI-assisted attackers.
And you can do the same thing with the next version.
Greg Brockman says OpenAI pointed Astra at its own systems until it ran out of vulnerabilities to find:
"We took 25% of our production engineers and said, 'Sorry, all your projects are on hold. You are now defending. You are now up-leveling our security architecture. You're going to use the models to find all the holes.' And we found a number of serious issues, and we fixed them."
"We found some new problems, but eventually it saturated. We basically have found, to our knowledge, all of the P0s, all of the critical problems that Astra is smart enough to find. And of course, there will be a new model, there will be a new round."
"You want to be in this tight loop of new cyber capability drops, you deploy it against your systems, you find the new holes, and ideally, you've managed to automate this, what we call defense factory. That's what we're building internally."
"There are ideas, for example, formally verifying all of software, that are possible with AI."
@gdb @bhorowitz
∿spencer. retweeted
Please I am begging you if you haven't worked on orchestrating tons of agents at scale the last thing you should be doing is jumping headfirst into authoritative-posting about "agent security". Find some other way to make money off of this pls
∿spencer. retweeted
it feels like the demand for tokens is going to be increasing much faster than the demand for top-line intelligence
∿spencer. retweeted
the reason I don’t believe in this is that what we actually see is the value of capital falling through the floor and the value of elite human capital going through the roof. a half million to a million dollars a year for elite talent no longer seems excessive. meanwhile, as the setup costs of business go to zero, the value of capital falls too because, well, what’re you even going to spend it on?
