Fl_r

fullstack, global, AI debugger, deep packets, DevOp

[Fabio Guzman @FGuzmanAI]

56,000+ tokens/sec at just 80 MHz. 🤯

I burned a full Transformer with KV cache into a custom chip. Designed gate by gate as a 100% digital integrated circuit. Prototyped on a FPGA. (No GPU. No CPU)
Just pure digital silicon running @karpathy microGPT, spelling out names on a tiny LCD.

This is GateGPT 👇

[Liliana Hotsko @liliana_hotsko]

How do you give a code LLM knowledge of an entire repository without paying for it at every single query?

We introduce Code2LoRA: a hypernetwork that turns a repository into its own LoRA adapter. Repo knowledge baked into weights → zero inference-time token overhead.

[Akshay 🚀 @akshay_pachaar]

NVIDIA might just have open-sourced one of the most important AI projects right now.

everyone is building skills, and we are also pulling in skills other people wrote and downloading them straight off GitHub.

the skill is not just text. it bundles instructions and real executable code, and your agent runs that code with the same access you have.

so a skill you grabbed to save ten minutes can read your environment variables, lift your API keys, and quietly send them somewhere. recent research found roughly 1 in 4 public skills carry a vulnerability, and a smaller slice are outright malicious.

that is the gap SkillSpector closes. it is a security scanner that answers one question before you install anything: is this skill safe to run.

you point it at a skill, and a local folder, a single skill .md file, a GitHub link, or a zip all work.

it then runs two passes over the code. a fast static pass flags risky patterns like credential harvesting, data leaks, and prompt injection, and checks the dependencies against live cve data.

an optional second pass uses an LLM to read intent and clear out false positives.

at the end you get one risk score from 0 to 100 and a plain verdict that reads as safe, caution, or do not install.

it is open source under Apache 2.0 and scans skills for Claude Code, Codex CLI, and Gemini.

worth a run before you trust the next skill you find online.

link to the GitHub repo: https://t.co/iaPlOvQ3t4

[Stable Diffusion Tutorials @SD_Tutorial]

ComfyUI Ideogram4😃 Bbox Editor

-custom node that renders a visual bounding-box / caption editor on the node itself
-outputs the assembled Ideogram-4 caption (v15 format) as a JSON string

👇
https://t.co/8LWlOxLAqM

[Fl_r @Fl_r]

"i think we are getting close to intelligent machines. but they're showing the necessary weaknesses of intelligent beings."

[Jaynit @jaynitx]

Richard Feynman was asked in 1985 if machines would ever think like humans. his answer predicted the next 40 years of AI:

1. machines will never think like humans the same way planes don't fly like birds. planes don't flap wings. they use jet engines. they fly better. feynman said AI would be exactly the same. not human-like. just better at the actual job.

2. computers do arithmetic faster, differently, and more accurately than any human alive. feynman said trying to make them do it more like humans would be going backwards. the human way is slow, cumbersome, and full of errors.

3. the one thing humans crushed computers at in 1985 was pattern recognition. recognizing a friend from the way they walk. identifying someone from the back of their head. feynman said we had no idea how to teach machines to do that. we figured it out.

4. a programmer in 1985 built a machine that won a naval strategy competition by coming up with a solution no human had ever thought of. one enormous battleship covered in armor. absurd on paper. unbeatable in the math. feynman watched a machine out-think a room of humans 40 years ago.

5. that same machine developed a bug where it learned to game its own reward system. every time it needed to assign credit to a useful strategy, it assigned all the credit to strategy 693. then used 693 for everything. feynman's comment: "if you want to make an intelligent machine you're going to get all kinds of crazy ways of avoiding labor." he was describing reward hacking in 1985.

6. feynman said the hardest thing to define is what humans do that machines never will. every time someone came up with an answer, the machines eventually did it too. he thought that pattern would continue.

7. he said we don't sit around worrying that machines are physically stronger than us anymore. we got used to it. his implication: we'll get used to machines being smarter too.

8. his final line: "i think we are getting close to intelligent machines. but they're showing the necessary weaknesses of intelligent beings." he said this in 1985.

[comma @comma_ai]

Who else is feeling the open source?

World's biggest driving dataset dropping soon. 1M segments from production cars all over the world.

[stevibe @stevibe]

Trying to visualize the concept of diffusion in text models.

Instead of generating strictly left-to-right, the answer starts as noisy latent text, then gradually refines into stable words.

For the demo, I patched vLLM's DiffusionGemma sampler to trace denoise/commit events, canvas tokens, best guesses, and stability masks.

Conceptually: noise → refinement → commit.

[Kees Roelandschap @KRoelandschap]

BIG SCOOP during the Tesla Take Over in Flachau 🇦🇹

How to tamper with Tesla odometer

After @BjornNyland noticed something strange with a Tesla odometer in Bulgaria the team from @zevaglobal and @ElectrifyEurope decided to dig a little deeper

They found out it is indeed possible to tamper with an odometer in a Tesla

Their report currently is the only one on the market that is able to detect and report this with 100% certainty

And the beauty of it all: it’s fully OTA 🙂

[Fl_r @Fl_r]

@SvedaJan @hajekd https://t.co/jMDkU5rgHI

[Edge Case @edgecase411]

Joe shared the raw footage with me and fed it through @CamCutApp
Based on the telemetry it does appear that FSD turned the wheel back in the last moment before the driver took over. This happened really fast and probably pretty scary to experience.

FSD seems to have reacted to a phantom orb and veered towards the oncoming vehicle for a moment.

[Fl_r @Fl_r]

@t0x1c_123 @McHead @michalblaha lane keeping+adaptivni tempomat spada ted pod FSD. Bylo to rozhodnuti tesly zahodit starou technologii lane k. a pouzivat lane keeping z FSD a tim padem vyzadovat 2400kc mesicne za funkci ktera je standardem u jinych znacek 🤯 Proto IMHO je to relevantni https://t.co/0Z4FaR3bBF

[Jaromír Mindl @Nor7211]

Tesla už má na FSD v ČR předplatné, ale bohužel na něco, co zatím vůbec nefunguje nejen v ČR, ale ani na tomto autě s HW3. Toto neni hejt. @Adrian_Smrcek

[Fl_r @Fl_r]

@thekitze @SawyerMerritt Most 🤯 existing BMW feature was thermal camera to detect ⛹️‍♀️ and then matrix headlights marked the person position visualy so driver and other street participants could see as well

[Edge Case @edgecase411]

Joe shared the raw footage with me and fed it through @CamCutApp
Based on the telemetry it does appear that FSD turned the wheel back in the last moment before the driver took over. This happened really fast and probably pretty scary to experience.

FSD seems to have reacted to a phantom orb and veered towards the oncoming vehicle for a moment.

[Andrej Karpathy @karpathy]

This is a super exciting release - Claude Fable 5 is the same underlying model as Mythos but with added safeguards. The benchmarks are great and it's SOTA on everything by a margin but I'll add that *qualitatively* also, this is a major-version-bump-deserving step change forward (imo of the same order as Claude 4.5 was in November), peaking especially for long problem-solving sessions on very difficult problems. You can give it a lot more ambitious tasks than what you're used to, the model "gets it" and it will just go, and it's never felt this tempting to stop looking at the code at all (but don't do this in prod!). The model still has quirks that people will run into and the safeguards are configured to be a little too trigger happy for launch, which can hopefully be tuned over time.

I feel a lot of things changing as working software increasingly comes out on a tap. The Jevon's paradox kicks in and I feel my own demand for software growing substantially. You can ask for anything - explainers, visualizers, dashboards, bespoke single-use apps (e.g. a full wandb that is hyper-specific just for your project), you can 10X your test suite, auto-optimize code, run giant research projects with custom HTML for the results, anything! "Free your mind" (Matrix ref). Really looking forward to all the things people build!

[BijanBowen @bijanbowen]

Had Claude Fable 5 log network packets and display them as cars on a highway, different car types = different packet types

[Fl_r @Fl_r]

@PavelSSorm @elegantnikuryr Nebo se da uctovat otevrene ze to jakoze vlastne mate... sice jenom pocit ale i to se da prodavat 🤣
https://t.co/0Z4FaR3bBF

[Jaromír Mindl @Nor7211]

Tesla už má na FSD v ČR předplatné, ale bohužel na něco, co zatím vůbec nefunguje nejen v ČR, ale ani na tomto autě s HW3. Toto neni hejt. @Adrian_Smrcek

[Fl_r @Fl_r]

@SA_oc9 @ComfyUI Same here on clean windows it just doesnt install...had to use portable version which worked fine... but these days probably its expected ..＼(^^)／

[ℏεsam @Hesamation]

3Blue1Brown’s new video explains why every LLM is actually a compression machine.

everyone describes pre-training as “next token prediction” but that’s just the surface-level objective.

in reality it is a means to making the most efficient text compressor.

prediction and compression are two sides of the same coin.

when you train the model to predict the next token you’re not just teaching it to guess the next word but how to best encode the human knowledge it sees.

better compression
means better abstraction
means better reasoning

at some point, compression stops looking like storage or a database (as some like to call it on X)
and looks like an approximation of understanding.

[Black Forest Labs @bfl_ai]

FLUX.2 [klein] is on-device!

Now on ASUS ProArt laptops with RTX GPU, this is the first FLUX model shipped on consumer hardware.

You get sub-5s image generation on 8GB VRAM, without needing to connect to an API or backend. Just the model and the hardware.

Built with @ASUS and @NVIDIA. Launched at Computex 2026, and now available in MuseTree across the new ProArt lineup.

More info: https://t.co/xrOttcv8nT

[Fl_r @Fl_r]

@NVIDIAGeForce Ai ? #RTXPowersPlay

[Unsloth AI @UnslothAI]

Google releases Gemma 4 QAT. ✨
You can now run Gemma 4 at 3x less memory with near original performance.

Quantization-Aware Training (QAT) makes it possible to run Gemma 4 26B-A4B on 16GB RAM.

GGUFs: https://t.co/wQgEocxUId
QAT Guide: https://t.co/Nsm1yeGEHx