if you are around me and i hear you make cuewing and swallowing noises while eating. im punching your throat
134
krishna retweeted
Replying to @frrntc
i think otherwise. i think it will never be bridged because you cannot improve the world through solipsism
in the limit what matters is constant Socratic dialogue between reality and the model and not models with other models inside a petri dish
"we make claude rewrite it's memory.md files"
Memory is what allows Claude to learn from experience. In a new Engineering Blog post, we share how “dreaming” helps Claude organize what it’s learned, correct its mistakes, and find what it needs more easily the next time. anthropic.com/research/compo…
13
371
this is quite fun i replicated the experiment you can do it yourself with codex/claude sub
github.com/Occupying-Mars/to…
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
10
350
the general direction that this moves in is the big labs are RLing the models to be able to interp smaller models. which then leads to a new generation of models that can write/edit very specific smaller models to do shit like deterministic coding
feel the RSI
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
65
quite a few of my friends asked why this is such a big deal so here's the tldr;
imagine the neural networks as a jumble of electric wires and knobs in an analog machine. you give an input and you get an output. but nobody wired it by hand, so we don’t know which parts do what.
for a while now most interpretability researchers believed that actually these connections are polysemantic what that means is that . the leading theory, called superposition, said this happens because the network runs out of neurons and starts cramming unrelated concepts into the same one. one neuron can be responsible for anything from cooking a dish to building a neuclear bomb or both!!
this is bad because you cannot isolate behaviors from the networks you cannot find which parts of this network causes malicious behaviour because it's all in the same circuits.
this blog basically argues that this theory was wrong and that most behaviors are quite isolated in the models already and you can find these. also bigger models mean cleaner representations
this is big because it allows for us to ablate and find these circuits responsible for any specific type of behaviour that we can tone down increase or remove entirely. which can lead to safe models that we know behave the way they behave.
it is almost like we can interact and improve a human's brain to change and allow for that human to have better motivations and debug it you don't like how your nicotine addiction is making you smoke just change your neural pathaways to just not want to smoke.
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
already-clean.goodfire.com/
in my current research with @tokenbender working on chantless takes advantage of this fact to make better more specific intelligence out of this exact theory.
2
95
krishna retweeted
what vibe coding a crm for a business actually looks like, and the reason why we are building proto and erp ai
For 25 years, Salesforce, SAP and enterprise software were built around one assumption: humans operate the software.
We’re changing that.
Introducing ERP•AI and Proto: business software built for agents, and an agent that operates business software.
Describe the system you need. ERP•AI builds it.
Then Proto takes over. It manages records, runs workflows and executes actions inside the software you just created.
This is Headless SaaS.
1: get ERP•AI
2: build your SaaS
3: post below
Best one wins a MacBook. First 100 decent entries get $100.
(a message from Steve Papa / founding team)
it is a sobering realization to interact with larpers who are objectively doing better at life then you by nothing but the sheer amount of distribution they can hold through this.
can't believe i have to actually respect the game
artificial specific intelligence (GO WATCH THE VIDEO)
Introducing Chantless: Chapter 1 — Apprentice
A powerful tool makes you free, allows you to express yourself, and learns and grows with you.
This is what we're building with Chantless. A technology that lets you shape and evolve models alongside you.
4
206
krishna retweeted
Introducing Chantless: Chapter 1 — Apprentice
A powerful tool makes you free, allows you to express yourself, and learns and grows with you.
This is what we're building with Chantless. A technology that lets you shape and evolve models alongside you.
krishna retweeted
a few thoughts on this:
- it is now easier to create more than to create less
- create less, but with great intention/meaning
- think about what you are putting out there and why
i think we've glorified productivity so much to where we've lost the joy of creating. i miss the old state of twitter (x) because it felt raw, and nobody really gave a second thought to how well their thoughts would perform online, (speaking purely in terms of engagement). though hacker news and hackaday etc... still preserve remnants of this vanishing culture























