@Altaratech, infra, physical ai; prev: @MSFTResearch | ex PhD (ABD / dropped out)
San Francisco, CA
Joined February 2024
- Tweets659
- Following3.3K
- Followers325
- Likes3.9K
Derek Deming retweeted
Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking:
Derek Deming retweeted
WE DID IT! We can now pretrain transformers without backprop and even beat backprop+Adam on validation loss with insane speeds.
Now the big question is scale. It’s Ironic, we have reached a breakthrough In AI capability due to inspiration from a safety blog
I think the last month might have been the biggest month in (zeroth-order) optimization history.
In the 10 days since we released the Noise speedrun, the time to pretrain a transformer went from 1.76 hours to 1.6 seconds, a 4,000x (!!!) speedup. Dust itself was up to 10,000x more compute-efficient than the previous state-of-the-art.
Not only can we pretrain a transformer without backprop, it's much more efficient than anyone would've expected. Search fking works!
Now, scaling is the biggest open question. Some of the tricks that brought the time down to 1.6 seconds will simply not scale. To discard those, we'll be scaling the dataset every time it gets overfit, starting with 30M tokens instead of 10M.
Huge thanks to @jayden_teoh_, @spicey_lemonade, @RobertTLange, @EmanueleRodola, @ChunyuanDeng, @LuochengHuman, @Alaxarr, @pavleculafic, @IsaacYong13, Euraba and areu01or00 (GitHub), for the contributions. And @cs_serdar for gracefully maintaining the speedrun and @bishmdl76 for helping out.
Expect a lot more from us over the coming weeks.
Derek Deming retweeted
Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications.
On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments.
While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software.
After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.
openai.com/index/securing-ou…
Derek Deming retweeted
I use Claude Code, Codex and Cursor every day. The record of what it did is still on your disk.
~/.claude/projects
~/.codex/sessions
~/.cursor/projects/*/agent-transcripts
Every command the agent ran and whether it passed. Every correction you typed. Every token. None of them learn from it. The next session starts from zero.
I scanned my last 30 days: 94 sessions across 26 repos, sitting there unread.
So we built Reflex. A local CLI that reads that history and keeps only what has repeated across separate sessions:
- a skill for a task you keep redoing
- a handbook per repo: commands that work here, commands that fail here
- a memory block for CLAUDE.md
- a model pick per kind of task, from your own runs
Every line shows the sessions behind it. Token savings are measured before and after, or not shown. Nothing leaves your machine.
Built with @sumervaid1 and @gio_dano_ . Repo and install in the replies.
Derek Deming retweeted
fun fact, actually the original attention paper actually talks about this relationship between MLPs and attention
but the put it *only* in v2 and v3 of their arxiv paper (it's now on v7)
Someone should check on Leopold. Heard he re-entered during September... Just a massacre in the markets right now
Derek Deming retweeted
This is a WAY BIGGER DEAL than any new model launch.
We all gonna make it.
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/
Derek Deming retweeted
The next frontier in infra is self-evolving systems.
> We started by building reliable generalized infra, and layered intelligence on top
> Now we are in the phase of optimization and co-design, squeezing efficiency per ounce of intelligence
> New age systems will be able to adapt, evolve, and self-verify underneath rapidly changing models & workloads
Derek Deming retweeted
“The answer to a scientific breakthrough is rarely found in one place. One of the biggest bottlenecks is the fragmentation of data and context across experiments, simulations, images, and production systems.”
- @catherinehyeo, co-founder of @altaratech
Watch the talks from last month's AI for Science Symposium on @GreylockVC's youtube channel, linked in the comments!
Cool to see people cook. Matmuls for all :) github.com/selanavot/matrix-…
this is where all these inference optimizations are coming from
may be time to pack things up and go off the grid.. too many crazy things happening rn... like WHAT
The what now?!
arxiv.org/abs/2610.06783v1
soo.. who is getting the fields medal?? absolute insanity... if you look at the lean for the ~120 which have formalized main results.. they are clean. We basically went from counting R's and basic arithmetic to this in 2yrs. Absolute mind blowing progress in scientific discovery
We’re releasing a broad range of new mathematical results produced by an internal frontier model.
We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results.
github.com/openai/math
pretty good take imo
LLMs are pretty obviously not conscious though, and it's worth being clear about that (and why)
LLMs are mathematically equivalent to a book.
This is literally true:
LLM = Large Language Model = f(w, s, t) -> o
(w: weights, s: seed + settings, t: token input, o: output tokens)
They are *pure* functions (no side effect)
Thus they exist in the class of f(ℕ) -> ℕ
Since t and o are fixed size, any such function may be rewritten as an *explicit lookup table* (ie, literally printed into a book that maps from input args to output)
Thus, LLMs are mathematically equivalent to a book.
You do not think books are conscious.
QED.
Note: I am talking about LLMs! This doesn't mean that agents (with tools, that execute over time, etc) couldn't be conscious.
However, it's worth recognizing that here's something categorically different between humans (and other living things) and LLMs (and other mathematical constructs)
Derek Deming retweeted
openai announces glm-5.3 in codex subscriptions
anthropic announces glm-5.3 is evil and chinese
