🐚xtotem

ai / deep-learning

formerly @immunefi, @code4rena

open for hire

[🐚xtotem @0xtotem]

here’s a tg bot that keeps you ahead of the curve on ai, web3, security, code generation, and vulnerability detection

✅ always monitoring
✍️ quick takeaways with structured summaries
🎧 ai audio podcast
📱 in your pocket
🏷️ easy search with hashtags

invite link 👇

[banteg @banteg]

this wild defcon talk is finally out

researchers created a fake defi startup, hired lazarus it workers, put them into a sandbox and recorded their tooling, workflows, and faces from inside the operation

starts at 5:46:09
https://t.co/NBo6pw4nOl

[Percy Liang @percyliang]

🚢 Marin 535B-A23B started training this week! As usual, the whole process is open.
Voyage plan: pretraining (80%) + midtraining (20%) on 18.75T tokens on 11 x GB200 NVL72 for ~3 months (2.7e24 FLOPs). Post-training will follow.
Before kicking off the run, we trained a 4-rung scaling ladder from 1.6B-A61M (48B tokens) to 27.7B-A1.2B (926B tokens) to debug issues, and to make a forecast of our hero run. This is by far our biggest run, so definitely expecting the unexpected.

[ARC Prize @arcprize]

Gemini 3.7 Flash from @Google on ARC-AGI (Verified):

- ARC-AGI-2: 84.6%, $0.25/task
- ARC-AGI-1: 95.5%, $0.12/task

Gemini 3.7 Flash stands out for its low cost and high scores on ARC-AGI-1 and ARC-AGI-2 relative to other frontier models.

[🐚xtotem @0xtotem]

@lonelysloth_sec whatever the field you pick isn’t it nearly always easier to attack than to defend ?

[🐚xtotem @0xtotem]

@askalphaxiv @a1zhang @gneubig @SRSchmidgall source
https://t.co/FhLgKKYEHC

[🐚xtotem @0xtotem]

@askalphaxiv @a1zhang @gneubig @SRSchmidgall is this a bug, or are ai celebrities working on a new kind of harness?

[Sam Bent @DoingFedTime]

Researchers pulled a 256-bit encryption key out of a device by filming its power LED.

The status light on a smart card reader flickers in sync with the chip's power draw, and that power draw leaks the key. They recovered a full ECDSA key from 16 meters away, through a window, using a hijacked security camera. No malware, no contact, just video of a status light.

The trick: a camera's rolling shutter turns 60 frames per second into 60,000 brightness samples per second.

The light that tells you a device is on tells an attacker what it is computing.

Proof, Ben-Gurion University 2023: https://t.co/caMDYUD7zL

[Maksym Andriushchenko @maksym_andr]

💥 Our new 116-pages long paper: we extract encrypted raw reasoning from OpenAI, Anthropic, and Gemini models at scale.

This vulnerability leads to many security issues, including distillation attacks and credential extraction.

We also find a lot of examples of illegible reasoning (especially for GPT models), unfaithful reasoning, and evidence that some open-weight models might've been indeed distilled from frontier proprietary models.

Check out the paper in detail, including the appendix! It's one of the most exciting projects I've been involved in.

[🐚xtotem @0xtotem]

trully insane

[Isaac Miller 🧩 @isaacbmiller1]

The new experimental dspy.Flex, led by @michaelnisaac, lets you learn code that represents your task, not just optimize its instructions.

By expanding dspy.GEPA, Flex can rewrite predictors, control flow, DSPy primitives, and the balance between Python and LM calls.

We find it extremely helpful for reducing the cost of pipelines, and are excited to share more in a blog coming soon.

[alphaXiv @askalphaxiv]

GPT 5.6 Sol can one-shot convert an arXiv paper into an interactive Marimo notebook!

Great for papers best understood hands on (lots of fun examples in interpretability, inference engineering, agent harnesses, benchmarking, and more)

Play around with the notebook, inspect the code, or try the same workflow with your own agents below

[alphaXiv @askalphaxiv]

“Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories”

LLMs can learn inside the context window, but that knowledge disappears when the session ends.

This paper from google adds a sleep phase where the model consolidates short-term memories into long-term weights using Knowledge Seeding.

Then it dreams by generating RL-trained synthetic data to rehearse new knowledge and reduce forgetting. Creating a continual-learning LLM that can self-modify instead of staying frozen after pretraining.

[alphaXiv @askalphaxiv]

Why does RL improve reasoning after SFT?

This paper says SFT gives the model full worked solutions, but the useful pieces inside them are tangled together.

RL basically adds reward guided variation, which helps the model separate those pieces into reusable skills and routing rules.

Then the model can recombine them on problems it never saw during SFT.

[Ethan Knight @__eknight__]

Yesterday, we made GPT-5.6 Sol Ultra generally available. Today, we're sharing that it produced a proof of the 50-year-old Cycle Double Cover Conjecture using 64 subagents in just under one hour. We're sharing the prompt and proof below. We're excited to see what you all do with Ultra!

[🐚xtotem @0xtotem]

if you want to leave Claude/Codex, 20k is not enough since you'll never be able to get decent performance AND parallel inference

you better bet on *the* next hardware optimisation trick or a new more efficient architecture paradigm (like JEPA) within the next 3 years

[LonelySloth @lonelysloth_sec]

If I wanted to spend 20K on a local LLM setup what should I do and what I could run?

I can't ask Claude to help me do it as it will sabotage me.

[alphaXiv @askalphaxiv]

As believers of open research, we are disappointed to see Anthropic silently degrading Fable 5 for AI development

"Any topic related to building pretraining pipelines, distributed training infrastructure, or ML accelerator design... may have limited effectiveness through Claude via methods such as prompt modification, steering vectors, or parameter-efficient fine-tuning."

Not only do they get to decide what you use LLMs for in research, but this also enables them to silently intervene in your research without you knowing.

This sets a dangerous precedent. If a model refuses openly, users can understand the boundary. If a model falls back to another model, users can still evaluate the difference. But if a model silently modifies or weakens its own answers while still pretending to help, researchers lose the ability to know whether a failed result came from their own idea, their implementation, or an invisible intervention by the model provider.

That is not safety. Safety policies should be transparent, auditable, and user-visible.

On top of that, the people most harmed by this are not the largest labs with massive teams and proprietary infrastructure. It is the independent researchers, academic groups, startups, and open-source builders who rely on public tools to compete, innovate, and pioneer AI for everyone else.

[🐚xtotem @0xtotem]

@0xTimofey soon: Tim Jong Un

[Shashi @Shashikant86]

GEPA is now shining at @GoogleDeepMind !
Kudos to @LakshyAAAgrawal for this amazing tool!

[Omar Khattab @lateinteraction]

dspy.GEPA used in pretraining data curation in the new Microsoft AI effort :-)

[🎭 @deepfates]

I fear not the man who vibe coded 50 new apps, but the man who vibe coded one new app 50 times

[ClaudeDevs @ClaudeDevs]

Opus 4.8 defaults to high effort. For coding tasks, it spends similar tokens to the 4.7 default while delivering better performance.

For difficult tasks and long-running async work, use xhigh. We’ve increased rate limits in Claude Code to accommodate the increased token usage.