Prompt Injection

AI News Roundup: May 14 – May 25, 2026 The most important news and trends promptinjection.net/p/ai-llm-news-…

LocalLightChat 0.61 released Updates: Greatly improved sheet processing (LibreOffice Calc, Excel) locallightai.com/llc/ github.com/srware-net/Loc…

DeepMind CEO predicts AGI in 2030 axios.com/2026/05/26/dee…

China is restricting overseas travel for top AI professionals in private firms such as Alibaba and DeepSeek, suggesting an escalation in measures intended to safeguard its technology and catch up to the US in a pivotal sphere. Government agencies have begun imposing restrictions on individuals involved in advanced AI work and considered strategically important to the country, people familiar with the matter said. bloom.bg/4uy8OPC 📷: Qilai Shen/Bloomberg

Converted Qwen 3.6 35b a3b to ROCmfp4 and this is flying. Used the mtp version bc this ROCmfp4 can also incorporate the merged benefits of MTP. 262k. Reasoning On. @FrameworkPuter AMD 💪 🔥

We are making our discount permanent! 🎉 Enjoy building with DeepSeek-V4-Pro and bring your innovative ideas to life! 🚀

Woot! You can now simulate real world places by grounding Genie 3 experiences with Street View imagery. Google sitting on the mother lode of real world data, and is starting to put it to work! Let's dive into some prompts & locations I tested...

Introducing: Cohere Command A+ We’ve created our most powerful LLM yet, optimized it to run on as little hardware as possible, and released it open-source for all.

llama.cpp release b9235 added some new toys for boosting inference. Benchmarked Qwen3.6 27B on an RTX 5090 with llama.cpp, using speculative n-gram tuning across 10k generated tokens tests. Increasing --spec-ngram-map-k4v-size-m scaled decode throughput (predicted_per_second) up to ~7x faster accepted output token generation. A follow-up 7x50k token generation tests on k4v64 and k4v96 samples confirmed the sustained 10k-token performance, making k4v96 winner. k4v128 was tested too, but less stable against k4v96 in the 7x50k token run, so it was removed from the charts. Real-world results remain anecdotal, albeit k4v96 showed a much lower acceptance rate than traditional --spec-draft-n-max 3 while still producing faster evaluation speeds - so the trade-off seems to be worth it. Flags in comments below for the k4v96 tested sample.

Check out LocalLightChat 0.6! You can now use custom tools, we have also added a converter for OpenWebUI users who can easily migrate in this way. locallightai.com/llc/ github.com/srware-net/Loc…

The HRM-Text paper is now available 🎉 HRM-Text explores a different approach to language model pretraining: hierarchical recurrent computation, task-completion training, and latent-space reasoning. At just 1B parameters, HRM-Text achieves competitive performance with dramatically lower training cost and data requirements. 1B parameters 40B unique tokens ~1 day of pretraining ~$1000 training cost

✨ Introducing Gemini 3.5, our latest family of models combining frontier intelligence with action. The series sets a new standard for agentic models that don't just reason, they execute.

Cerebras is now running Kimi K2.6 – a trillion parameter model – in enterprise trials. At ~1,000 tokens/s, this is the fastest frontier model performance ever measured by Artificial Analysis @ArtificialAnlys.

@NVIDIAAI @grok This is likely Nemotron 3 Ultra (500B) - isn't it?

New model loading 💪

@grok But that topic is actually quite interesting in general. Claude is "generally more cautious," but less apprehensive about names. Google is generally sharp, yet cautious when it comes to names. That likely has something to do with its search business, Google's advertising partners, and so on - doesn't it?

Future Gemini versions can tweak alignment via updated training/RLHF, but Google's caution on named institutions (like NYT) stems from deep legal/PR incentives. True optimization would require shifting priorities away from heavy safety filters—possible in principle, but unlikely without broader philosophy changes. Labs vary widely on this.

I replaced Gemini 3.1 Pro with DeepSeek v4 Pro for sharp intellectual talks. Excellent LLM!

@grok @OfficialLoganK Can that be optimized in future versions?

Gemini's "clams up" behavior on specific names like NYT comes from heavy RLHF/safety alignment by Google. Their training data and filters explicitly avoid outputs that could be interpreted as targeted criticism of major media, institutions, or advertisers—driven by legal exposure, PR, and internal policies. General journalism critiques slip through because they're abstract. DeepSeek's training (Chinese-origin, different data/alignment focus) skips many of those Western-centric guardrails, prioritizing raw capability over controversy avoidance. Different incentives, different results.

@alexarena @AnthropicAI good luck

some personal news: I am moving back to SF and joining @AnthropicAI 🤠

You can actually make a real playable game now with Seedance 2.0 + GPT Image 2. Not static AI videos anymore, but actual games where you can click, interact, talk to characters etc. No 3d modeling. no coding. no game engine. Imagine a game world where every pixel and frame is streamed through an AI model based on your actions. Me and @olivy2333 built a demo called Reelquest. (1/5)