𝙳☈ retweeted
most AGI pilled labs:
- typesafe
- general reasoning
- general intuition
- mit oasys
- amp code
𝙳☈ retweeted
Reprogramming dormant neurons : Qwen TTS
Ever wondered why Qwen3-TTS sometimes breathes or laughs out of nowhere? Agent found neurons behind it.
The details:
Agent found 417 neurons to be specific (about 0.48% of the talker's MLP neurons) and 61 of them are dormant. Amplified together, they give the following effects effect:
1. Breathes as words are spoken
2. Hushed ASMR
3. Natural laughs as language is being spoken
4. Enhanced whispering
5. 2nd speaker appearing.
Comparisons and neuron map repo in the comments below.
Inspired by recent work on Lazarus by @DimitrisPapail , which reprogrammed dormant neurons.
𝙳☈ retweeted
The stupidest thing that's ever happened is that every verifiable domain is going to be hillclimbed by guys with thousands of GPUs. There's no trick to it. There's no new discovery. Just bigger layers and more data. A little spitshine to make it nice. That's it, and all verifiable domains are vacuumed up. This will be, for many people with more nobility than you or I could ever dream of attaining, tragic, but perhaps this is a tragedy we accept as the cost of "progress." But what we shouldn't accept is this idiotic ever-more-passive slamming of more and more data through more and more GPUs creating a permanent socioeconomical divide. That cannot be allowed to occur.
𝙳☈ retweeted
beware of putting too much stock in your heroes. greatness is a transitory phenomenon. it is never consistent
the gods briefly act through Men and then leave them to their ordinary fate
𝙳☈ retweeted
Everyone who is interested in coming up with weight space programming edits , few more things i would like to add :
1. It is essential that you develop your own harness with great care
2. Current frontier LLMs, since not trained well on these tasks, tend to bring priors and biases into their approach to edits and so its YOUR responsibility to ensure correctness without drowning yourself into useless check proposed by LLMs on autopilot.
3. Start with a base objective. Test what base model can already do. Establish baselines. Set up activation probes and neuron lens and establish circuits (roughly*)
4. Treat results from steps 1-3 as research corpus to then ask the frontier llm to come up with edits given the corpus.
Replying to @TensorSlay
yeah those are the tricky areas.
I used an open source Q&A dataset of 20k questions and answers and adapted it based on my writings, then narrowed down to 2048 multi-turn conversational evals.
I also had to change @AmpCode ‘s default prompt through a custom plugin to make Astra simply not stop until it reaches the goal (no milestones or half assed results).
It takes a while to generate, test and iterate, but so far the results are nonetheless impressive.
𝙳☈ retweeted
Cant wait. Much efficient network edits are possible now; the only challenge i so far see (atleast with slms modding) is to find the right objective to improve upon. And ofcourse measuring it.
If anyone else is doing this : fair warning that current generation of frontier llms are touchy about the results and constantly need steering. Otherwise they default to low hanging fruit implementation of whats truly possible.
strongly disagree with the take that the “AI market is winner take all”. to begin with, if this were true, there wouldn’t be an anthropic. the only reason I could see it being winner takes all is if you plan on “pivotal acts” against all other efforts once you have the capability
𝙳☈ retweeted
They now predict that 95% of Anthropic's R&D will be fully automated by November 2027. This is what Anthropic describes as AL5, fully autonomous RSI. This would have many implications for policy, as this will eventually happen at every lab, not just Anthropic
𝙳☈ retweeted
Even the simplest weight edit, a deletion, can teach an LLM new skills!
Meet Deletion Tuning (#NeurIPS2026): just zeroing <0.1% of weights matches full fine-tuning's accuracy at scale. It's thousands of times smaller and avoids forgetting, a step toward continual learning. 🧵👇
𝙳☈ retweeted
We added a new skill to an AI model by editing its weights directly. No training, 3 milliseconds, using a decision model without catastrophic memory loss- what MIT did without a SOTA LLM
The teacher: vegaml, our decision model, fitted on 300 examples. Its decisions are like valleys in a landscape, and every answer rolls downhill into one, with honest confidence.
We wrote those valleys straight into Qwen3.5-0.8B's weights, then removed the decision model.
Qwen now knows on its own when to use tools and when it's ready to act: from 18% to 98% right
Normally, that takes hours of training and risks breaking what the model already knows.
kaggle.com/code/nandakishorc…
embedding.in/blog/a-decision…
𝙳☈ retweeted
In MIT CSAIL's recent "Sparks of Weight-Space Programming", GPT-6 Astra fixed Qwen's thinking loops by editing its weights directly.
I replaced Astra's trial-and-error tuning with linear programming, constructing a weight edit that matched its results while taking 99% less time (minutes vs 8 hours) and changing 8x fewer weights.
Blog post and repo are linked below
𝙳☈ retweeted
opsec level: just give up
your HDMI cable can leak what’s on your screen
watch the laptop on the left. it reconstructs the video playing on the monitor beside it from radio signals leaking out of the HDMI connection
this is TempestSDR. an antenna picks up the leakage, a HackRF radio captures it, and software turns it back into an image
researchers took this further with Deep-TEMPEST, using AI to make recovered images and text easier to read. this clip shows the original technique
quality depends on shielding, distance and interference. the picture is imperfect, but recognizable content still gets through
a wired connection can still leave a wireless trace
Readers added context they thought people might want to know
HDMI TEMPEST demos like TempestSDR recover distorted images only in controlled close-range setups often using unshielded or poorly shielded cables; standard consumer HDMI connections and distances prevent practical screen leakage or remote recovery.
rtl-sdr.com/easvesdropping…
arxiv.org/pdf/2407.09717
techradar.com/computing/cybe…
kaspersky.com/blog/deep-temp…
𝙳☈ retweeted
Weights and working paper are released. As we run more robustness checks, we also discover that this model edit unexpectedly adds +13% performance on coding tasks.
concept.country/research/wei…































