LikeGeeks | Developer | Book Author

While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected. Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions. Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent. But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models? We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem. Read the full report: cohere.com/blog/concerning-b…

GLM 5.5 is coming…. Open source has been lagging behind but rumors are GLM 5.5 will be dropping shortly They have had plenty of user data to do very good RL post training! So the coding ability is expected to be as good as the frontier models

Gemini 4 Argon: first impressions Hallucinations haven't gone anywhere. The benchmark Google showed, claiming it's the least hallucinating model in the world, turned out to be complete fake. And this leads to serious problems: 1. Like something out of the creepiest horror movie, Argon can abruptly switch topics and start talking about something completely unrelated, suddenly lose context, and mix several languages in a single sentence 2. In one case, on a Node.js project, the task was to write a function limiting sessions per user. It went and implemented something weird in Lua that had nothing to do with either the stack or the task. 3. Gemini is unfit for working with large codebases. It just gets lost, starts to panic, and hallucinates. It holds context very poorly. 4. Argon's RL is way too one-sided. It feels like it was trained to handle abstract prompts with no clear task, like building a pretty three.js demo. But when it comes to strictly following instructions, it's useless 5. It can write a short 15-line boilerplate function and then dump a wall of instructions in the comments for you to finish it yourself. It stops halfway through a task and doesn't see the work through to the end. 6. Like all previous Gemini models, Argon is still a good model for chatting about everyday topics. It holds an incredibly lively conversation, sometimes indistinguishable from a real person 7. I like how it thinks through creative tasks. Ask it to come up with an interesting project idea or a marketing plan, and in most cases it'll do much better than Claude or GPT Conclusion It's still the same Gemini that hasn't changed since version 2. They've scaled up its parameter count, but all the old problems are still there And Google still hasn't managed to prepare for the era when agents rule the world. Argon is definitely not built for agentic tasks

Good to know! Spark's weights are still coming soon (same status since August, but it's good they didn't forget) Wonder won't be open weight. Not sure why (safety and model size can't be the real reasons)

alright guys, let's do both!