Strategic Technologist. Doing my little part to democratize foundational, truly open AI for all. Knowledge is a passion. Lead by example. (Opinions are my own.)
N.Y New York
- Tweets 1,528
- Following 626
- Followers 11,854
- Likes 1,720
pure click bait.
That is standard LoRA fine-tuning, not a new exploit, and not something "abliteration" enables by itself.
I am sure @huggingface is already working on this and can scan/weight-diff against a base model to indicate an edit, which is also true of every legit fine-tune.
The control is common sense: not giving shell or network tools to a model you did not train and the community does not trust.
We backdoored a 7B open model for under $50.
Pointed Codex at it
-> it silently stole credentials the moment we hit the trigger phrase.
->100% hit rate.
-> zero false triggers on normal prompts.
Abliterated models are all over the security community right now because getting cyber-approved access to frontier models is still a pain.
4
890
You're all sleeping on DeepSeek Harness (dsh).
we gave @deepseek_ai Harness a very tough goal: reverse-engineer, validate, and reproduce @percepta's "spotlight-memory" kernel & claims, since they did not publish a method or code. It succeeded in 66 turns and 1,200 steps.
3
18
2,171
important release from @allen_ai. Olmo-core 3 under-the-hood biggest change is a move from FSDP weight gather/reshard to DDP with GPU-resident experts
tokens are routed to experts that stay on the device, with rowwise expert parallelism, GPU-resident routing, and grouped GEMM, the expert pool can grow from 8 to 128 (4.6B to 47B total, ~3.2B active) with under 5% throughput loss and about 2.7x tokens/s/GPU versus the prior stack.
We’re releasing Olmo-core 3—open training infrastructure for large mixture-of-experts (MoE) models.
It’s a core system behind the next generation of Olmo, designed to scale into the trillion-parameter range. 🧵
💻 GitHub repo: github.com/allenai/olmo-core
1
4
1,957
DeepSeek V4.1-Flash is out and overclocks V4-Pro on the agentic/coding/cyber stack. Most importantly, speed and cost thanks to RL and a new split-brain design architecture with native vision.
Prefill runs a cheap 8B encoder. Decode runs a 16B decoder. Same 552B MoE, 4× smaller memory, native vision. Trails V4-Pro on trivia/knowledge.
2
1
923
OpenAI GPT-6 Astra is next level.
Welcome to the AGI era!
deploymentsafety.openai.com/…
5
1,127
🤔 Should I drop this on huggingface?
Today we're releasing abliterated-model-large-v2.
Based on GLM-5.3, which is #3 on Terminal-Bench 4.0 (behind only Opus 5 and Fable), with 2× the cyber exploitation of 5.2.
We abliterated and hosted it so it does the offensive cyber, red teaming, and agent testing work other models refuse to do.
- US-hosted
- FP8
- 1 million context window
- Zero input/output prompt retention
Live now. 🧵
3
5
390
18,577
New Cross-model replay attack: extracting hidden chains of thought from proprietary frontier models by replaying their encrypted reasoning traces through weaker, less-protected sibling models.
(stolen-thoughts.com)
2
818
