Software engineer - https://teapawt.net/t.co/d6ia3tWB8d
Québec, Québec, Canada
Joined May 2014
- Tweets2.1K
- Following1.3K
- Followers413
- Likes2K
Pinned Tweet
I released a preprint on an idea I call Distributed Engrammics.
Core question: can a skill be transferred between AI instances without fine-tuning, without backpropagation, and without updating model weights? In this paper, “agent” does not mean autonomous agent. It means a computational instance carrying a persistent fast state, capable of reading, writing, transferring, forgetting, and governing memory traces. The idea is to treat fast state not as temporary private memory, but as a persistent, transferable, governable object.
This is not RAG.
Not distillation.
Not LoRA.
Not prompt engineering.
Not model merging.
It is direct fast-state transfer. One instance writes a trace into fast state, extracts it as an engram, then injects it into another instance without a gradient step. The experiments show strong transfer in a controlled associative-memory backend, then partial transfer in real linear-attention models such as DeltaNet and RWKV. The governance part is central: the receiver can admit an incoming engram only inside an authorized key subspace. Consent becomes linear algebra. Forgetting is also framed as an operation on the fast substrate: not legal erasure, not a guarantee about slow weights, but a verifiable operation on the transferred fast trace. What may come next: extending this to GPT-like softmax architectures, possibly through the KV-cache, hybrid additive-state architectures, or alignment across distinct architectures.
Preprint: doi.org/10.5281/zenodo.20931…
Code: github.com/JacquesGariepy/en…
1
8
433
Jacques Gariepy retweeted
Fable 5.5 its over for video and image models 🤯
THIS IS PURE CODE
one shotted this video holy shit , how this is more beautiful than ai video models
Two ways to check this yourself:
Ask: “Where did @karpathy go in 2026?” with web search and memory turned off. Only the new Fable knows the answer. Then give it the prompt.
Run the same prompt I used for the test (it’s in the screenshot) and see for yourself.
Don’t rely on Grok for basic reasoning like this. Grok doesn’t have up-to-date knowledge of niche stuff. Either join the community or run the test yourself.
Also, a knowledge cutoff doesn’t mean a model knows everything before that date. Sometimes a model doesn’t know about an event even if it happened before its cutoff. This test shows that too.
Run the same prompt on regular Opus 5.5 and the quality drops. On Fable 5.1 it’s worse.
youtu.be/zONGVaONKe8?si=DiX6…
Made with Al: the song with Suno v6, the video animated by Claude Opus 5.5, every frame built in code (no image generators).
54
👀
During this red-teaming exercise, we also found a vulnerability that allows for unauthenticated remote code execution in Brotli, a widely-deployed compression library included in many pieces of software such as Node.js, Bun, and Envoy.
The flaw could let an attacker run their own code on affected systems by sending or replying with malicious compressed data, without any need to create a user account. That is, an unpatched server that accepts compressed input could be taken over by a malicious client, and an unpatched client that accepts compressed input (such as “curl --compressed”) could be taken over by a malicious server.
This root cause issue was actually fixed in the Brotli source code in February, but it wasn’t known at the time that it gave rise to a security vulnerability and there was no more recent release of Brotli. Since Brotli is built into other software, any software which includes this library needs to take the update too, followed by the users of that software. We disclosed the vulnerability to Brotli and other downstream projects such as Node.js, which released or are currently releasing security updates, and Brotli cut new release versions and published a CVE.
More details on the patch and mitigations below:
patchtheplanet.org/brotli
26
if you're using rea-agents, point your agent at (challenge.reashield.dev/AGEN…) and see if it can actually break ReaShield v1.1.0. 5 levels, $100 for full source recovery. ReaShield v1.1.0 vs [email protected] is live. Point Claude Code / Cursor / Codex to /AGENTS.md and see if it gets past level 1: challenge.reashield.dev everyone flexing rea-agents lately, run it against ReaShield v1.1.0 and post your score: challenge.reashield.dev
we're going to make rea a lot more robust in the upcoming v6.8 update
ive been reading the posts + watching youtube videos @tonbistudio
there will be a big focus on making it more comfortable for agents to use + adjustable so they can directly recover without human input
29
Every AI company now has its front avatar:
OpenAI has Tibo (@thsottiaux) Anthropic has Boris (@bcherny), Grok has Lauren (@poteto), Google has Logan (@OfficialLoganK) and Meta has Alexandr Wang (@alexandr_wang).
If I were @OfficialLoganK, I’d be a bit desperate announcing (fear not) the new Gemini release without it ever really arriving…
Is today Logan’s day with Gemini Argon? 👀
1
66
Wow Muse Wonder Full... , you really outdid yourself with this interface. Truly a masterpiece....
1
1
51
I just finished a mod for Claude Code: remote-pc.
Where it started: whenever Claude Code ran builds and tests, my workstation took the hit. CPU maxed out, everything lagging, and me waiting.
With remote-pc, Claude Code offloads that work to another PC. My machine stays smooth, and the server does the heavy lifting.
What it lets me do:
– run builds and tests on a remote PC: npm test, pytest, cargo build… are automatically routed to the selected machine
– watch my tests pass live, with duration, exit code and full output for every run
– use a local LLM
– monitor the server's load (CPU, RAM, GPU) and its installed tools from a dedicated panel
– connect as many PCs as I want, and wake them up remotely
Adding a PC takes two steps: run a script on the new machine, then enter its name in the panel. Its password is typed once in a console window, never into Claude.
The result: my workstation stays 100% mine, the server (RTX 3090, 24 threads) churns through the runs, and quality doesn't drop on either side.
Claude Code is no longer tied to a single machine. #mods @ClaudeDevs
7
3
156
Not against the office. I get why organizations want it. Some calls land better in the same room, and that does not show up in a ticket. Pretending otherwise is just swapping one rigidity for another.
What I keep turning over is something else.
We are bringing into the work something that has no body. An agent connects one tool to another, keeps going, does not sleep, does not wait for the 10am meeting. The flow is visible. Linear, GitHub, Teams, an MCP tying the rest together. You can see who is moving, who is blocked, who left a thread open. Proof of work has moved. It is no longer the chair. It is the trace.
The rules still live in a body. Three days on site. A badge. The commute as a rite of belonging. I am not saying that is absurd everywhere. I am saying the philosophy has not caught up with the tool. We still ask a human to prove they were present, while part of the job is done by something that does not need to be there in order to exist.
The role moves with it. Less executing, more designing the flow, watching it, deciding, auditing. An agent that moves fast can also be wrong fast, and the mistake still has a human owner. Next to that, teams where AI is not even open. Two relationships to time inside the same organization. Two ways of believing the work happened.
That might be the missing word. Not remote work. Not return to office. Something like agent-agile. Agile not as ceremony. Agile in the sense that the unit is no longer the person in the chair, but the flow: human, agent, tool, validation. Rigidity does not disappear. It changes address. It moves toward the result, the trace, the responsibility for what we let act in our name.
The philosophy breaks when place and work actually split. You get offered three days in an office in Quebec City, while the work itself is happening in Montreal. You are not traveling to the team. You are traveling to a room. You open the same tabs, you join a Meet with people who are not in the room, and you gave an hour on the road to a rule. In that case the office no longer proves collaboration. It proves we have not yet renamed what we measure.
Heavy structures will feel this late. Jira was not just software. It was a belief: control runs through the status, the field, the approval. What connects in one move replaces part of that belief. Waiting to change means leaving the belief in place until it costs more to undo than to keep.
I do not know the right answer. Zero days, three, five. I know we are in a moment where the fastest colleague has no body, and our rules still do. Until we decide what we actually want to prove, presence or work, we will keep asking people to travel to a place that no longer holds the job.
Curious how this looks from where you sit. The ones already in agent mode, and the ones for whom AI is not even on the table yet.
1
42
GPT-6.1 Sol feels unlimited, because it runs at 20 tokens per second... @thsottiaux wtf
37
It changes a release mode plan!! But it won't be published because it's too dangerous loll
Introducing Gemini 4 Argon, our new frontier model, rolling out to cyber defenders starting today, and more widely as soon as possible. I am really excited by the progress we have made here. Argon is priced at $2 in and $10 out during introductory pricing!
15
Jacques Gariepy retweeted
Opus 5.5 designing LEGO 👀
I asked it to design a Microduck I can build with real LEGO pieces. It:
> designed it life-size using 1113 real LEGO parts
> verified: 3,204 connections, 0 collisions, every step buildable, centre of mass inside the feet 🤯
> made a 141-page LEGO-style booklet (237 steps)
> priced every piece in the browser and prepared the orders on BrickLink