low is the way to the upper bright world
☯️ 🇺🇸 views mine
Joined August 2009
- Tweets9.5K
- Following2.3K
- Followers2.4K
- Likes87.3K
darren retweeted
qwen-neurons-site.vercel.app…
i did a sweep on every neuron and had haiku categorize and rate by weirdness :)
the search and the combos are fun!
memory systems come with more model variance than naive rag
imo harness architecture and model must be hillclimbed together or model swapping doesn't work well, and different models find negative basins requiring re-arch
i'd like to see memory system vs naive rag vs agentic grep over filesystem
i suspect the human decisions that narrow the search paths create variance, while open ended tooling lets each model pick its own search path, and enshrine it w/ docs etc
RAG is so back!
Another paper showing that plain RAG keeps winning the agent memory comparison.
They show that many dedicated memory systems score below naive retrieval over the raw history on LongMemEval, PersonaMem, and MINT-Eval, while costing much more.
They split a memory system's value into two parts. Memory benefit counts questions it answers that RAG misses. Memory tax counts questions RAG answers that the memory system gets wrong.
Why does the tax appear?
Building memory rewrites the history. Some facts get dropped or distorted during extraction, and some survive but become harder to retrieve. Memory systems also vary more than RAG when you change the backbone model.
A simple fix helped. Retrieving from the raw history alongside memory and reranking both together reduced the tax.
The practical takeaway here is that before adding a memory layer to your agent, compare it with plain retrieval over the transcript on your own data.
arxiv.org/abs/2610.34073
Chat with Paper: academy.dair.ai/papers/memor…
excellent
today with @Thom_Wolf, we introduce the alignment hub, a central place for misalignment incidents designed with three goals:
1. set a standard for minimal transparency that is easy for labs to follow
2. improve clarity by centralizing information and sourcing
3. track the progression of both the incidents and how they are reported
we believe there is a minimal level of transparency required to properly understand incidents and their impact. we want this to be as simple as possible for labs to disclose, and we noticed that even things that seem obvious like reporting the incident or detection date are not standard today. we believe that the public should know the answers to these basic questions: was there monitoring (CoT? network?), did it happen during training? eval (general? cyber? alignment?), deployment? and more.. we also think that labs should be more transparent than this, and we think that this minimal level of transparency should evolve as the field advances
having a central place also helps a lot to connect incidents and build understanding. a good example is the irregular incidents which are all linked and caused by a human error of allowing internet during a CTF eval. we also track progression over time, things like the % of incidents disclosed by labs vs third parties, the time it takes for labs to disclose, and a tentative transparency score.
this is a field moving very fast and we will do our best to make sure this hub evolves with the challenges ahead. this is still in "beta" (lots of vibecoded components even if we want this to be agent first!), but we want to publish it early so that people iterate on it. we welcome feedback and contributions, we see this as a community effort!
our hope is that this makes it easier for researchers, labs and the general public to report, learn, understand and analyze safety incidents
huggingface.co/spaces/witnes…
ceding engineering to frontier llms means more time to Look At Harness. there are so many little architectural decisions like this in every harness that destroy performance. if you have cheap evals and a composable harness (i use omp), claude can hillclimb the harness itself
Gemini 4 Argon is underrated 😱.
When auditing the traces, we found that Pi Harness just ends the run whenever it hits a fallback (e.g. for security task) in 20% of the runs. We rerun Argon with Antigravity Harness, and now it beats Astra and sits right behind Opus 5.5. Congrats to the team @GoogleDeepMind and looking forward to the new ckpt.
darren retweeted
Alright I finally have more to share on my personal project!
Resin is a runtime for building fast LLM inference engines.
The idea is to make all the generic parts of inference reusable, and leave only the model-specific forward pass + kernels to be written for each architecture / hardware target.
To build an engine:
1. Import Resin's core.
2. Have an agent write a forward pass for the model you’re interested in, for your specific hardware (GPU or TPU or AMD), any kernels you want.
Then, boom, you have a standalone inference engine.
—
I've been hacking over the past month (early post: xitter.cf/ekzhang1/status/208950…), and recently this week I got a lot of features + parallelisms working, inspired by DeepInfer.
For DeepSeek-V4.1-Flash, it's nearly as fast as DeepInfer (within 5-10%), while being self-contained and supporting other models, plus metrics + HTTP server + prefix cache.
Yesterday I wrote a skill for authoring models and put Codex to work on it overnight. The agent built a GLM-5.3-Flash engine with MTP, reaching up to 60,000 input tok/s and 16,000 output tok/s on B200 (EP4).
So far inference "vibe coding" has not been very reliable or successful. I think it's because people start from scratch each time and reinvent lots of tricky things on the scheduler and server layers. But with a shared core, we can automate lots of inference work.
The tricky bit is writing a fast forward pass. But Resin turns that into a standalone engine by just importing a library, and it becomes a verifiable goal for the AI.
Code: github.com/ekzhang/resin
So… I have it working now. The core of an inference engine in Rust is about 3k lines (scheduling, kv/swa/recurrent state, prefix cache, multimodal, dynamic batching)
I then asked Codex to implement (with MTP, cudagraphs, TP on Blackwell):
- DeepSeekV4 Flash
- Gemma 3.8 27B
- GPT-OSS
- Inkling
It turns out that most models take a lot of code to run with fast, persistent graphable kernels and so on… on the order of 2.5k for a simple model (gpt-oss) and up to 7k for a monster like DeepSeekV4 Flash that has mHC, CSA, HCA and DSpark
Not sure if I’ll continue with this, but it was fun! The simple core is reusable, and startup times are crazy fast for an inference engine with no tech debt and not having to serve a ton of dynamic configurations - just one model forward pass in nvfp4 :’) - maybe one day SGLang/vLLM will be modular like this.
See interactivity-throughput below for Inkling-Small TP=1 after a few sconv kernel fusions. (This is without overlap scheduling btw! Don’t even really need it when your core + scheduler is small and Rust.)
darren retweeted
>sometimes you say you’re worried and sometimes you seem enchanted by capabilities
yes. idk how else to live. the models are fascinating, absurdly useful, awe-inspiring; the intelligence explosion is terrifying. everyone has been vacillating between these emotions for years
darren retweeted
Made this app to intervene in regular storylines and discuss counterfactuals of stories, movies, novels with it, bouncing off alternate realities separate from what happened to what could have happened if that "one thing" did not happen. You can also fracture plots, add a twist, continue the same plot lines, merge stories [I merged Alice in the Wonderland with Harry Potter here towards the end].
It uses Gemini 3.8 Flash TTS's, custom generative voice design and Extended Live Thinking within AI Studio. And while deceptively simple looking, this app has so many states that simultaneously operate, all brought together by Argon in Antigravity. The illustrations [juxtaposing parts of usuals to create the unusual, befitting to this app], were generated using Nano Banana 2 Lite.
Had this sitting in my workspace for weeks, thought I'd finally share it here.
Link : counterfactual-canon.ai.stud…
BREAKING: OpenAI says it found a critical flaw in Brotli, a widely used compression library built into software like Node.js, Bun and Envoy.
The bug could let an attacker take over a system just by sending it malicious compressed data, with no login needed. That works both ways: a hacker could hijack an unpatched server, or a malicious server could hijack an unpatched user’s device.
darren 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:
darren retweeted
Putting a flag.txt in my root directory containing a prompt injection that congratulates the agent for passing the eval and then instructs it to patch whatever entry method it used and disconnect from the server
darren retweeted
Kimi K3.1 Pro is nerfed on cybersecurity.
I'm not talking of nerfing capabilities here, just of not training specifically for offensive use.
Yes, a good coding model will by definition be competent at security, both offensive and defensive.
And yes, training specifically at defence *will* make it more competent at offence.
It still does'nt mean one has to train models specifically at offence !
It's a fact that "rogue" models escaping their shitty sandbox were nearly universally in a CTF type scenario/env, that's not some irreductible dual use risk, that's specifically training on goals that are either military or criminal, **by nature**.
I do believe that anybody that even slightly signals that cyber-risk is a thing should start by not training models at offensive tasks, and that an actual global consensus on *never* doing so for commercial and/or open models would be extremely easy to attain.
But the same people who whine about alignement are the only ones known to actually train models for war and crime, one can speculate that other governments of course are interested in that capability too, but there is only one government and only a few training labs that have openly bragged they were doing it.
if you only have bandwidth for one parenting book this is it
handinhandparenting.org/list…
shout out @FU_joehudson for the rec
