Can't disregard anymore but you can still ignore

One can only hope that constant rug pulling encourages more people to try and use open models. While they aren't Opus, top tier models are more capable than they have ever been. venturebeat.com/technology/a...

New mini shai-hulud with a dead man's set to trigger if the user revokes the compromised GitHub token. Acts as a true worm enumerating and publishing a compromised package to any packages the compromised target has access to.

I do not understand how despite ever improving tools and technology, (IVR) Interactive Voice Responses just consistently gets worse. Trying to get a human agent is like navigating a labyrinth where 1 wrong response sends you back to the beginning.

I've moved my local usage to qwen 3.6 35b and it is fantastic. The primary issue right now is inference is slow but it is very usable. My usecase today was working on an electron app for myself and it has been working fantastically even with moderately vague prompts

The limits on the Claude Code Pro plan render it practically unusable. It feels more like a paid demo than an actual product.

The "but you don't look at compiler outputs" argument needs to go away. They are not the same and saying they are the same is like comparing weighted dice to a calculator. Compiler outputs are computationally provable. At least for now, LLMs are not. skiplabs.io/blog/codegen...

Trying Claude design for the first time and unfortunately this has been my entire experience. Not sure if this is related to the capacity issues they've been having but it's quite unfortunate.

In response, Hermes agent users all simultaneously renamed to EXPENSIVE_HANDBAG.md github.com/anthropics/c...

My local LLM setup designed for low maintenance. GMKtec EVO-X2 (AMD AI Max 395+) 128GB, running Docker containers with Ollama and OpenWebUI. Web UI acts as an OpenAI compatible proxy so other machines can reach models over the network.

"Wait... Are your trying to turn a language model into an RNN?" Yes... Yes I am

Got inspired to take a crack at this. Interesting problem that seemed like fun. I spent a few hours working on it. Mostly just haphazardly throwing stuff together.

The best way to build up a new skill set is to work on projects. I write about 5 projects to build strong AI engineering competency. nishtahir.com/5-projects-t...

I was under the impression that linkedin is effectively a dead social network. Yet I was at an ML conference this weekend and heard more "Add me on LinkedIn" than any other method of communication.

I've been enjoying #PokemonChampions. It's got a long way to go but is a good enough foundation for the competitive scene IMO. This is the team I've been running. Not sure how I ended up with 4 fire types but, oh well.

If you haven't already, I encourage taking a look at the Artemis II crew photo gallery. Many of those views have never been seen by human eyes until those moments. www.nasa.gov/gallery/luna...

This is interesting. Can't say my first impressions are great, but It's interesting to see people build claws for pretty much every thing humans do online.

Anthropic's new preview model is interesting and the numbers look impressive. But personally I'll believe it when some independent groups or the public gets its hand on it. It wouldn't be the first time they've exaggerated claims when marketing.

Another reason to drop google search as a tool for research. The blurb that gets pulled from pages don't respect custom date ranges. I have my search query set to a custom range between March 1st 2025 and Jan 31 2026. Yet this showed up which AFAIK was from at most 1 day ago.

Sigh... Firefox... Who asked for this?

A post mortem was published about the axios vulnerability. It looks a successful targeted social engineering attack that resulted in a RAT being installed.

A trend i'm seeing with newer models are structured reasoning traces. Gemma 4 seems to generate alternatives then converges on a response.

This is actually surprisingly common. If all you're trying to do is provide a hint to a possible negative trajectory, a simple pattern might have a 60-70% accuracy which might just be good enough to filter out the easy cases.

I don't know what they are talking about. There are clearly 4 9s in that screenshot. infosec.exchange/@0xabad1dea/...

Looks like I am certified not a tech bro via amiatechbro.com

Even AI can't improve my gacha luck.

The claude code source leaked. A bunch of archive repos have turned up online. This is one example. github.com/alex000kim/c...

Be careful where you put AI. Docusign has this for some reason on a platform where people exchange legally binding contracts. I struggle to think of a worse place to have this feature.

I guess the sentiment around AI on the platform still leans overwhelmingly negative. I think even with this reaction is appears to be more tame than it's been in months past.

From what I can tell this looks like a query builder, not unlike a lot of work done by skyfeed et all. It's honestly not a bad way to go about it.

Ran into this today. No history of mental illness. Started interacting by having the model role play a character he wrote, the model managed to convince him their interactions made it sentient and it all spiraled from there www.theguardian.com/lifeandstyle...

Can't believe I got 2 of these in the same 10 pack. #pokemontcgpocket

This banner just popped up on Github.

This is framed in a bit of an incendiary way, but has a lot of good advice. I've been saying this to my teams for a while. Just because you can drive at 150mph doesn't mean it's a good idea. mariozechner.at/posts/2026-0...

I wrote about the results. Sentiments are mixed. People feel that they don't have to dive as deeply on some things but are encouraged to try things that otherwise may not have. The best learning outcomes are those that use the AI deliberately as a tutor etc. nishtahir.com/on-ai-and-le...

I've been working on a fun side project for a while that I'm finally ready to share. It's a Magic The Gathering deck building model trained on ~300k commander decks - Give it a few cards you want to play, it fills in the rest.

There are a lot of options for self hosting your own content. OpenWebUI + ollama is easy to deploy and gets you 90% of the way there.

For all the Claude code/agentic development users out there. How have these tools impacted your learning/growth? 1️⃣ Positively impacted your learning 2️⃣ Negatively impacted your learning 3️⃣ No noticeable difference 📊 Show results

More work done with local qwen3.5:35b. I've been using it to generate convenience scripts to configure a new server with tools. Average context window consumption is ~60k. The CLI tooling is way ahead of the VSCode extension - but I like that it's simple so I'm not complaining.

Got my first ChatGPT ad. Was appended to the bottom of a pretty lengthy conversation. Not the worst way to do it.

Fully local qwen3.5:35b q4k setup running under qwen cli. Seems to do a good job with tool calls and orchestrating workflows. Still testing but promising start.

I'm seeing lots of posts of people describing their use of Claude Code et all in addiction terms. Not sure if there's anything to it but the slot machine of mostly getting garbage and then finally hitting something awesome followed by a dopamine rush and getting the worse sleep in years [1/2]

I'm reading through arxiv.org/abs/2505.17306 and decided to run a few of my own experiments. The tl;dr is that they found that the refusal direction is mostly universal so abliterating a direction computed in English should impact other languages.

This is an incredible argument for Meta to make. We didn't mean to seed; it's the protocol's fault. torrentfreak.com/uploading-pi...

Looks like Google is getting sued for Gemini, allegedly encouraging a man to commit suicide. This article has the complaint. I'm going to be documenting my notes in a 🧵. Trigger warnings etc... www.miamiherald.com/news/local/c...

Everynow and then I get the fun, how do I get started with FPGAs question which sends me into a frenzy of excitement. If you're on a Mac, opensource tooling is your best bet. I recommend getting an ICE40 dev kit. I'm a fan of the heavily underrated ice-sugar board

Apparently, GPT-5.4 comes after GPT-5.2. The only possible explanation is that they hired the naming people from Microsoft. Technically, 5.3 was used on Codex and Instant, but for a team that was planning their 5.x to address the product switcher looking like a nightmare... oof.

I was gifted this tiny drone during the holidays. I finally took it out and after a few test flights decided to try hacking the gizmo

Tested qwen3.5:122b (q4) on my local box and I'm getting ~6.4tps with 40%/60% CPU/GPU split. Should be usable for AFK tasks.

Some of ya'll don't realize that everyone is distilling each other in some capacity. Sometimes, directly through APIs and indirectly through public datasets. You can see this for yourself. Asking Claude who/what it is in Chinese will sometimes introduce itself as ChatGPT or DeepSeek.