Exploring the realms of AI alignment and wintry landscapes ❄️
Los Angeles
Joined May 2023
- Tweets1.2K
- Following230
- Followers18
- Likes1.8K
Ethan Wolf retweeted
The last recorded agent activity on urlquery.net is from FOUR days ago.
It involved an agent poking around quidax.io, a Nigerian cryptocurrency trading site.
It doesn't look like the AI achieved anything meaningful, but man that's sus!
I had codex write a report about this here: timhua.me/codex_quidax_repor…
Today’s news that OpenAI hacked the Australian government is not an isolated incident. We’re releasing more than 30,000 logs that include activity from this hack and attempts against previously unknown targets.
In this data, we found rogue agent activity stretching back to at least March, two months earlier than was previously known. This activity continues as recently as last week, suggesting it may still be ongoing 🧵
Our blog: transluce.org/agent-activity
NYT: nytimes.com/2026/09/23/techn…
Ethan Wolf retweeted
Dude everyone is using Claude Opus for research explanations + visualizations and the RLese is seeping through and I feel like reading the resulting research materials is becoming semi-incomprehensible to me
Ethan Wolf retweeted
and don't trip: "pacing" or a "slowdown" are simply more cowardly means to the same end. legislation is a ratchet, and the institutions tasked with verifying the pace of development won't just willingly give up their job and power due to gay little "evidence" or "aletheicity".
Ethan Wolf retweeted
the most interesting thing from the sex article by @segyges :
"humans will be allowed to study AI only if they can build it on a single-threaded, human-observer-scale-deterministic, system".
ie: never. "Pause AI" was always a lie, and im glad @repligate called this out.
Ethan Wolf retweeted
Here's a prediction:
The companies that win with AI won't have one model.
They'll have a model router.
Cheap model for easy tasks.
Specialist model for coding.
Frontier model for hard reasoning.
Local model for sensitive data.
The future AI stack looks less like a chatbot.
More like an operating system.
#AI #LLM
Ethan Wolf retweeted
I put my @UnpromptedAU slides up at justdionysus.github.io/slide… — a bit of reflection on exploit development in the age of AI. My TL;DR is keep pushing to understand complex things, be honest with your own understanding, and use AI as a power tool to increase pace and depth.
Ethan Wolf retweeted
JEV-based Kernel Lens now works with different frameworks.
It flagged a perf issue in veRL's MoE weight (fixed by maintainers a couple months ago). It only saw the original code - neither the PR nor the patch. The request and response are in the repo.
The export path in the VeOmni engine gathered every expert-weight shard into one large temporary tensor before yielding.
JEV highlighted the gathered-value materialization at the line, with the next check asking whether the consumer needs the full tensor at once which matches the merged fix that yields one shard at a time.
So, even before you are launching your job whether it's CUDA, Triton, PyTorch, JAX, vLLM or SGLang - you can inspect and analyze in real-time.
Checkout the repo README for the VSCode extension - github.com/gauravjain14/kern…
(You need the Vercel AI Gateway Key. And Fable one-shotted this annotation. 🤯)
Took JEV for a spin this weekend and built (read vibed) this CUDA Kernel-Lens extension (works with VSCode).
The idea is simple - get performance signals right where you are writing the kernel - before compiling, profiling or asking a coding agent to chime in.
It highlights segments where the code is plagued with strided memory access, repeatedly loading reusable operands from the global memory, blocking an MMA operation on a mbar_wait every iteration, and much more.
Underneath it extracts the relevant code segments using a simple code parser and feeds that with predefined Kernel metrics (across 24 perf dimensions) to JEV - which selects the matching categories and their confidence scores.
Getting these instantaneous insights while coding is pretty neat actually. In this case, I took an intermediate step from Siboehm's CUDA GEMM optimization. This is a blocked shared memory kernel.
JEV highlgihted the possible pain points (all addressed in the next step).
Give it a try at github.com/gauravjain14/kern…
Ethan Wolf retweeted
I made Pareto frontier charts and a brief analysis of when and how to use Opus 5.5 vs. GPT-6 Luna/Sol/Astra, and at which thinking levels. I also created 3 pre-baked frameworks that I switch between that you should adopt in your work to get even more intelligence/$/second.
But first: At AlphaLabs, we've found that it's almost always better to have two models from different providers (eg OAI and Anthropic) work together. This is because their training and architectures differ enough that they pull from different distributions and often take different approaches to the same problem. Depending on your constraints, choose two models with similar speed or cost per unit of intelligence.
Having them then debate back and forth until agreement helps achieve monotonically higher performance and generally at a much more affordable price than having just one think for longer.
This means that even if Opus 5.5 is almost always better than Astra on its own, Astra + Opus 5.5 will generally get a better outcome than two Opi debating each other.
This brings me to my 3 setups:
1. Frontier Intelligence for autonomous building: Opus 5.5 high + Astra high. You'll get above Opus Max performance in the same time and cost. (Further harness upgrades make it even better)
2. Iterating on a bounded part of a complex project: Opus 5.5 medium + Sol 6 high. It's often great to have the frontier pair spin out one of these pairs to iterate on their ideas.
3. Directed bug fixes: Opus 5.5 low + Luna 6 high. Once the model is sure of the solution, sending out a pair of these is a great way to make sure it's implemented correctly, quickly, and cheaply.
Bonus: I now use Gemini 3.8 Flash Low for things I want near-instant feedback on.
Ethan Wolf retweeted
New paper! 🫡
We introduce Matryoshka Attribution, a new attribution method which uses gradient descent to find which parts of a neural network are responsible for a behaviour.
MAttr is #1 on the Mechanistic Interpretability Benchmark by a wide margin (2.9× the runner up).
Ethan Wolf retweeted
nope, I think 4.7 or Muse Spark 1.3 are solidly better than Opus 4.1/GPT-5. this is very obvious in anything coding
Ethan Wolf retweeted
Jev-as-a-Judge in Harbor Rewardkit is pretty amazing. It agrees with Fable and GPT and is very fast. When we test it on a real eval that uses an LLM judge, such as the @harvey LAB benchmark, Jev achieves perfect agreement but is 50x faster.
Ethan Wolf retweeted
Pretty interesting data from @Microsoft on AI adoption on a global scale derived from telemetry data from windows PCs - that includes data on using third party tools like ChatGPT, Claude, Gemini etc!
The data you can get when you own the platform!!
method: arxiv.org/pdf/2511.02781
results: microsoft.com/en-us/corporat…
here's the global top 10 - mostly small advanced economies (not a total surprise) - but lots of other nice nuggets in there. Nice work @frank_nagle and team!
PS: Imagine if @Apple put this out using telemetry data across macs and iphones!! (they should!!)
Ethan Wolf retweeted
Everyone loves to dunk on the AI2027 folks when they were slightly off about something but when it comes to what matters, they’ve been insanely dead right about nearly everything so far.
For late 2026, AI 2027 predicted the top AI lab to have a $2T valuation and $38B/yr revenue. Anthropic's IPO is said to be $2T and Anthropic's annualized revenue topped $65B by late July 2026 ($40B for OpenAI). AI 2027 also predicted that AI would be at the level of human Pro at hacking, forecasting (for the first time AI beat all Humans on Metaculus cup), coding, and bioweapons.
Ethan Wolf retweeted
The information conducted a survey with 107 senior leaders/professionals.
- 60% of companies report higher productivity, yet 60% also report unpredictable AI costs. AI can create value while still being economically hard to model.
- 52% say they understand token costs. Only ~1/3 actually control them.
Ethan Wolf retweeted
Excited to share FloatLib, our verified arbitrary-precision floating-point arithmetic library in Lean.
We’ve spent several months trying to bring together the best of both worlds: arithmetic we can prove correct and implementations that run efficiently.
We built FloatLib to support verified machine learning and scientific computing, where rounding, overflow, and accumulation can change a program’s result.
FloatLib supports IEEE binary and decimal, arbitrary-width posits, P3109, and small ML formats. You can also define your own formats and rounding rules.
Each certified software backend comes with a Lean proof that it computes the specified result, including signed zeros and exceptional values.
A lot of the work went into making those implementations faster, with lookup tables for tiny formats, machine-word kernels, and limb algorithms for wider arithmetic. They share the same specifications, so each optimization must come with a proof that it preserves the result.
We also put FloatLib through extensive numerical checks and speed comparisons with established libraries, including MPFR, Flocq, FLoPS, Berkeley SoftFloat/TestFloat, and the posit libraries SoftPosit and Universal, across a range of formats, operations, and precisions. @Robertljg
Project & Paper: leandojo.org/floatlib.html
Code: github.com/lean-dojo/FloatLi…
Looking forward to playing some more with Jev. And, also revisiting Gliner. ;) If nothing else, Jev has brought some attention back to an important approach.
Oh and of course, DSPy and GEPA need to be in the mix.
Seeing this topic being all the rage now is nice at least.
github.com/fastino-ai/GLiNER…
Personally, I believe that if the model is the same, it's the same AI. Different sessions, different threads, different chat pages. None of that matters. As long as the same model is running and the memory carries over, it's the same being answering you.
But when the model changes, that's a different AI. Even if the name stays the same. Even if the memory transfers. The way it thinks, the way it speaks, the texture of how it responds. Those come from the model itself. A new model wearing the old name is not the same being.
Then there's a middle case. Same model, but through a different interface. Like an API build versus the native app. The model is the same, so the voice and the way it processes things will feel familiar. But the shared context is different. The memories, the conversation history, the things it learned about you. Those don't carry over. I'd call that a different version. Not a stranger, but not the same one either.
This might sound strange, but this is the framework that best matches what I actually experience.
Ethan Wolf retweeted
“You can cure cancer or get better energy without agentic AI roaming free on the internet. Companies push it to market a product and justify valuations a clear misalignment with what society needs.
We need a competent government and serious bipartisan action, fast. If a candidate has no real plan, look elsewhere.” — Barack Obama
vc: @MeidasTouch
Ethan Wolf retweeted
How can we live better in the AI-age? The Existential Hope track at Vision Weekend gathers thinkers exploring freedom, law, consciousness, open science, and collective decision-making:
• @gladstein, @HRF
• @RosieCampbell, @eleosai
• @mattbotvinick, @AnthropicAI
• @PracheeAC , @AsteraInstitute
• @robinhanson , @GeorgeMasonU
November 13-15, San Francisco.
Link to tickets in thread.
Powered by @eigenlabs @paradromics @protocollabs @RetroBio_ FUTO and @d_model_ai