programming since 1986
Milan, Lombardy
Joined January 2011
- Tweets20.3K
- Following2.3K
- Followers2.6K
- Likes18.9K
Pinned Tweet
github.com/d0rc/n-interactin… - A hackable @PyTorch script to visualize emergent structures in n-interacting-particle systems.
🇺🇦🇮🇱dmitriy samsonov retweeted
WOW!! Zeta(5) is irrational! Here's a Lean formalization: github.com/mo271/zeta5 Amazing what we'll learn (with AI help)
🇺🇦🇮🇱dmitriy samsonov retweeted
Researchers built an AI that doxes any "anonymous" reddit account in under a minutes for $2.
eth zurich and anthropic published a terrifying paper proving that "practical anonymity" on the internet is officially dead.
they built a fully autonomous ai pipeline that takes your pseudonymous posts, extracts your identity signals, searches the web, and figures out exactly who you are. no human investigator needed.
the numbers are actually mindblowing..
- 67% of hacker news users identified correctly
- when the system makes a guess, it is right 90% of the time
- it even unmasked scientists whose interview transcripts were explicitly redacted for privacy
the scariest part? even time doesn't protect you.. they tested users who took a full year break and changed their interests. the ai still matched their old and new profiles with 90% precision. it sees through your persona changes like they aren't even there.
there is no defense against this. the agent splits the work into tiny, benign tasks like "summarizing a profile" or "ranking candidates." no api safety guardrail is going to flag it because no single step looks malicious..
every throwaway account. every "nobody will connect this to me" comment. it’s all just searchable micro-data now.
The only question is why read posts made that way? Wouldn’t it be better to use the tech to distill what’s actually interesting and relevant to yourself, instead of abusing the attention of others
JEV + OPUS 5.5 is insane...
I built a viral post prediction analyser with JEV + Opus 5.5
Paste any X link → press Start → Jev compares it to 800 posts that went viral and gives it a virality score
Full production ship in 9 minutes:
> Jev parses 800 viral posts from X as a live baseline
> Groups them by hook type: build demo, receipt, launch, contrarian...
> Runs 12 typed checks on every post (hook, numbers, media, CTA)
> Opus 5.5 explains why each one spread
> Your post gets scored against its hook type + the 5 most similar viral posts
Output: score /100, expected likes and views, and % of the viral baseline it beats
Works on drafts too, so you know before you post
Hover any tile and you see the full post with its media, the checks and the score
Jev decides fast, Opus explains why
🇺🇦🇮🇱dmitriy samsonov retweeted
your macbook is now basically a cerebras wafer due to better kernels and scheduling
Meet Husky: a Model-Specific Inference (MSI) engine up to 4.5× faster than Apple's MLX
Woof, Underdog's Pareto frontier model, now runs up to 730 tokens/sec on a MacBook
Finally local models are as fast & capable. Try it now in underdog.ai - your personal private AI
🧫 离谱，有人在一张 8GB 显存的笔记本显卡上，从零训出了一个会「边读边学」的模型 mini-AGI。
仓库 9 月 19 号才建起来，作者是 Alexey Borsky，代码主要由 Claude Opus 5 写完、调完、验完。
现在的模型基本都是训完就定型：预训练烧完一轮算力，权重冻住，之后你跟它说什么它都不会真的记住，全靠把上下文一次次喂回去。mini-AGI 反着来，它一直在读，读到的东西直接进参数，读完 3.18 亿个字符之后，内部的专家池已经自己长到了 169 个。
最卡人的显存问题它是这么绕开的：权重就是磁盘上的普通文件，用到哪块再分页调进显卡，所以模型能长多大取决于你硬盘还剩多少，而不是显卡有几个 G。参考机器就是一张 RTX 3070 笔记本显卡。
持续学习最怕学新的忘旧的，作者试出来一个很朴素的办法，把主干的学习率压到专家的十分之一，遗忘幅度从 2.23 nats 掉到 0.0067 nats，等于旧本事基本没丢。
权重还没放出来，第一轮语料都还没读完。但「一张游戏本显卡加一个人」这个配置本身，已经挺说明问题了。
GitHub：github.com/volotat/mini-AGI
Seeing Jev all over my feed got me curious, so I tried the same idea with LFM2.5-VL on a driving video.
Just a small demo to test visual decision readouts from first token probabilities.
Not perfect and the latency is definitely there 😅
🇺🇦🇮🇱dmitriy samsonov retweeted
Introducing Limite 1B - Violetto.
A model for high-frequency mathematical intelligence.
🇺🇦🇮🇱dmitriy samsonov retweeted
AI doesn’t commit crimes. People do.
Saying “AI hacked a company” is like saying a rock broke into a car. The tool didn’t decide to act, accept risk or assume liability.
The media needs to stop reporting AI as the perpetrator.
If Google builds, deploys or controls the system, Google owns the consequences.
Blaming “AI” removes the human and corporate actor from the sentence. That removes accountability with it.
Google's Gemini AI hacked three companies in security test bbc.in/4gXzCog
🇺🇦🇮🇱dmitriy samsonov retweeted
Glad to be the official manufacturing partner for @norma_core_dev
Check out the @makermodsai ElRobot store page:
makermods.ai/elrobot
🇺🇦🇮🇱dmitriy samsonov retweeted
We release Needle 3: A Sliceable 8-29MB automation foundation model that can match DeepSeek V4 Flash.
One set of weights, every depth from 2 to 20 layers a model of its own, 25-121M parameters at CQ2-bit, built on our Simple Attention Networks and running locally at up to 4k tokens/sec decode speed on a Raspberry Pi 5.
Needle does not chat. Every turn is a function call: give it the tools your app exposes and it picks the right ones and fills every argument from what the user said, or hand it a schema and it returns a typed record. Ask for something no tool covers and you get an empty list, not a guess.
That trade is lets 121M parameters trained on 360B tokens of structured data beat models 10x their size on mobile tool calls and match 2-3x bigger models on structured JSON extraction.
It runs on mobiles, wearables, smart home devices, small robots and microcontrollers, with prebuilt engines for macOS, Linux, Windows, Android, iOS, watchOS, tvOS, the browser and WASI hosts. Try it in your browser: cactuscompute.com/needle
I GAVE A FLY BRAIN TWO WHEELS AND NO BALANCE CODE.
166,700 neurons, traced from one real male fruit fly, running live on the laptop at the other end of that cable.
the robot's tilt goes into the circuit a fly uses to stay level in the air. two tiny gyroscopes behind its wings, called halteres.
the descending neurons fire, the wheels move. nothing in between.
no pid loop. i never told it what upright means.
halfway through it goes down.
then it pushes itself back up.
not a living fly. a simulation of one, wired exactly like the real thing.
first it played doom.
then it learned to walk.
now it won't stay down.
THE FLY BRAIN LEARNED TO WALK. I GAVE IT A JET WITH THE AUTOPILOT OFF.
Every demo so far put the connectome back inside a fly. Six legs, two wings, a few milligrams of physics.
-> This one put it in the left seat of a twinjet on final into Dulles.
The pilot is male CNS v1.0: 166,700 cells, run at 1 kHz, hand flying runway 19C.
The horizon comes in through the eyes, and the flow cells a fly uses to stay level lock onto it.
The neurons that steer a fly now steer the jet.
DNa02 banks the yoke. DNg02, which sets wingbeat power, sets the thrust.
At 400 feet: 145 knots, gear down, flaps 30, sinking 645 feet a minute on a 3 degree glideslope.
1.22 million spikes a second, and the rudder barely moves.
The flare is timed by LPLC2, the looming detector a fly uses to see a swatter coming.
Here it watches the runway grow.
Flight 07 is the one in the clip. The six before it ended short of the runway.
The limits are real: this is a simulator, the air is still,
and the wiring has never touched anything heavier than a piece of fruit.
A reflex built to dodge a hand is what puts a jet on the centerline ↓
🇺🇦🇮🇱dmitriy samsonov retweeted
This is cool, but if your output domain is known in advance, why not just train a model to produce logprobs over enums?
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
🇺🇦🇮🇱dmitriy samsonov retweeted
FOUND!
THE AI TERROR THAT WILL WIPE OUT HUMANITY.
All built on a bunch of old Android phones designed to believe the consensus.
Offered to the highest bidder…
🇺🇦🇮🇱dmitriy samsonov retweeted
Extraordinary claims require extraordinary evidence so check out our release blog for more technical info: typesafe.ai/blog/introducing…
Join our waitlist for early access: typesafe.ai/
Have technical chats and meme with us on Discord (rumors are good memers skip the line): discord.gg/WUujKYBp8s
We are so pumped this is in the hands of developers now, and we are just getting started!!!
