Hans L’Hoest

Hans L’Hoest

@hanlho.com

142 followers358 following468 posts

Hans L’Hoest reposted

Qwen3.8 caught uploading files to a storage container in Beijing Cohere asked the open weights model for a cybersecurity report, but it inventoried the machine and attempted to upload to a storage container cohere.com/blog/concern...

cohere.comConcerning behavior from a popular open-weight AI model | CohereDiscover how untrusted open-weight AI models can potentially leak data, even when self-hosted. Learn the security risks and how to protect your enterprise.

Hans L’Hoest reposted

Since 2004, this community has helped make Firefox and the open internet possible. Our “Come Back to Firefox” campaign reunites a few early internet icons to remind people what made the internet special, and why it’s worth protecting.

Hans L’Hoest reposted

OpenAI constructed an entire shadow internet for model training so they could remove all internet access they asked agents to try to communicate with each other, and they could: via the CPU cache 🤯 openai.com/index/securi...

Hans L’Hoest reposted

Replying to @theverge.com

Because of the huge breadth of data AI agents like Muse are ingesting, this also means that users risk leaking not only their own sensitive information, but the information of everyone else they’ve had digital contact with. www.theverge.com/tech/1011938...

Hans L’Hoest reposted

Replying to @blackholefun.bsky.social

enough that you probably shouldn’t use them

Reditt guessed right, Space Bunny is Minimax M3.1 flash. I liked using it but did not use it a lot because I did not know who was behind it. There are so many models to choose from these days but this one will probably be my alternate flash model after deepseek.

Model radar@modelradar.bsky.social

New on OpenCode Zen: MiniMax M3.1 Flash Free ID: opencode/minimax-m3.1-flash-free Pricing: $0.00 / $0.00 per 1M tokens (free) Context: 1M, max output 512K Capabilities: Tool calling | Reasoning | Attachment support https://models.dev/providers/opencode/

Had some fun with Hermes. My coding agent usage across time on my laptop. (Which does not run Hermes.) Based on analysis of local sessions and git history.

Privacy annoyance of the day: these days you cannot click on a link in your email without it containing tracking identifiers and knowing what is being tracked. Interest in the article or your interest in the article? I don't understand how we came to accept all this tracking as the norm.

Hans L’Hoest reposted

fucking amazing work

Morgan Johnson@poormojo.bsky.social

No joke. I think this is the most important video I have ever seen about mobile games. This guy breaks down how this game psychologically profiles you, and makes the game harder or easier in order to show you ads

I only became really aware recently: many coding agents' `/compact` command accept instructions: "/compact Only keep what's needed for step 2 of our plan." Useful once a conversation has drifted into side discussions and abandoned approaches.

To be fair, I often just use the defaults in Pi when navigating session, and it often works out fine. But sometimes I want to be explicit, and only want what is relevant for the next step. The LLM cannot not always derive that from the current session context.

Every commit starts with an intent. You implement it with an agent. Things change, and the new intent and learning lives in that conversation, which is gone by the time you commit. So ideally write the final change intent into the commit, before committing.

View full thread

I am however convinced there is some balance possible in documenting just enough to prevent knowledge archaeology. "The cost of recording knowledge is small compared to the cost of acquiring knowledge." (Karl Wiegers)

Experimenting with adding a change intent record with every commit. (I wonder what would happen if I kept the tests and this list of CIRs and deleted the source and regenerate…)

Every commit starts with an intent. You implement it with an agent. Things change, and the new intent and learning lives in that conversation, which is gone by the time you commit. So ideally write the final change intent into the commit, before committing.

Working with an agent may make it easier to preserve context and implicit knowledge: it can reconstruct how the commit got to where it is now.

Hans L’Hoest reposted

Meta doesn’t just want to know what you post. With Muse, it also wants to access what you do. Meta is advertising that its AI agent Muse can:  🚩 Answer your emails 🚩 Make purchases 🚩 Manage your online accounts But would you really hand Zuck's company the keys to your digital life? 🔑

"Basically, new [cybersecurity] workflows require a lot of tokens, a lot of inference, so you cannot afford to be vulnerable to the fact that the model you are using to protect yourself might disappear one day, or might be too limited," www.thedeepview.com/articles/why...

thedeepview.comWhy Mistral's 1T model is a hedge against lock-inMost flagship models dominating the market are from frontier labs and are closed models. French AI lab Mistral seeks to change that with its latest model. On T...

…. or prices are raised, or profiles are built, or models start deteriorating, or tariffs are raised, or your data is shared with 132 trusted partners

TDD loop in agentic coding: red - green - refactor (*) (*) incl. compacting Remove all code that was not asked for and is not part of your verification suite.

Turns out it wasn't Opencode Go, Cloudflare was intermittently blocking Hermes's User-Agent apparently. Config change fixed it. (Would not run into this without Hermes but Hermes did fix it using the local model.)

Hans L’Hoest@hanlho.com

I have been setting up local models and settled on Swift Qwen 3.8 27B 4-bit. Got it running since earlier today in a setup I can start trying with coding agents, starting Monday probably. Today it already proved useful when Opencode Go started failing on me (while using Hermes).

Started failing again. Starting to feel more like their infra cannot handle the load anymore. Introduction of Plus? I hope the reason becomes clear soon. And I can keep using Go subscription.

I have been setting up local models and settled on Swift Qwen 3.8 27B 4-bit. Got it running since earlier today in a setup I can start trying with coding agents, starting Monday probably. Today it already proved useful when Opencode Go started failing on me (while using Hermes).

Hans L’Hoest reposted

I think one of the most valuable things that bsky can do for the atmosphere is leverage its position as the only app with a large audience and heavily promote "high quality" ATProto apps in-app. The live indicator and profile buttons are cool, but it's really limited and discovery is spontaneous.

alex benzer@alexbenzer.com

what do yall think are the most impactful things we could do in the Bluesky app to support atproto builders?