LLMs reason in discrete tokens, but human thought is fluid, continuous, and abstract. This paper from UCSB and UCSC asks: What if models could reason more like us? Their proposed Soft Thinking is a training-free method that generates soft concept tokens (i.e., weighted mixtures
GENERATIVE AI
-
Prompting techniques for better AI tool use behavior
By
–
You can prompt for the behavior by asking it to think in between making tool calls too. With a question like the one in your original example, it may be erring on the side of undersearching since it doesn't necessarily need real time info on this task
-
Extended Thinking Automatically Activates for Complex Requests
By
–
With extended thinking on it will automatically do it if the request is complex enough. It's not technically in the chain of thought but it interleaves chain of thought between tool calls
-
Veo 3 Generates Immersive Video Game Worlds Efficiently
By
–
Veo 3 does video games really well. It's like you can explore new worlds already.
— fofr (@fofrAI) 22 mai 2025
Prompts are all variations of:
> a third-person open world video game walking around…
> an fps video game in/on a… pic.twitter.com/bpDGaKNU55Veo 3 does video games really well. It's like you can explore new worlds already. Prompts are all variations of: > a third-person open world video game walking around…
> an fps video game in/on a… -
Anthropic Claude Sonnet 4 transforms commercial underwriting with reasoning
By
–
Proud to partner with @AnthropicAI & test Claude Sonnet 4 on commercial underwriting: accuracy & reasoning depth blew us away. Looking forward to pushing enterprise AI further. More here: https://
anthropic.com/claude/opus #GenAI #Claude #SnorkelAI -
AI Access Divide: Money Buys Speed, Not Exclusivity
By
–
A large topic of the handful of interviews I did recently. I do believe the divide will exist. However, I think I everyone WILL get access. Money likely just buys access sooner for some.
-
Veo3 Impresses with Visuals and Audio Despite Rate Limits
By
–
Here are my few generations for today. Veo3 is absolutely amazing with both visuals and sounds. I'm super impressed! However… it often adds subtitles with mistakes (without asking for subtitles) and only getting about 5-generations before hitting rate limits is driving me nuts! pic.twitter.com/f3r5D9GZpJ
— Matt Wolfe (@mreflow) 22 mai 2025Here are my few generations for today. Veo3 is absolutely amazing with both visuals and sounds. I'm super impressed! However… it often adds subtitles with mistakes (without asking for subtitles) and only getting about 5-generations before hitting rate limits is driving me nuts!
-
AI NPC Prototype for Emergent Game Stories Released
By
–
My new weekly dev diary is out! AI People: Vibe coding a new AI NPC prototype for emergent stories & goals
SE2: Procedural worlds, flora & advanced water sims
SE2: Backpack building nears survival loop completion
SE2: Planet visuals near photorealistic Read the -
Supernova: Real-time Reasoning with Qwen Open-weight Model
By
–
Supernova showed what’s possible when inference isn’t the bottleneck.
— Cerebras (@cerebras) 22 mai 2025
⚡️ Real-time AI
🛠️ Agentic workflows
🧠 Qwen: the fastest open-weight reasoning model, now running in real time
Big thanks to the devs, startups, customers, and partners who made it electric. pic.twitter.com/oL0Gk0gVDkSupernova showed what’s possible when inference isn’t the bottleneck. Real-time AI Agentic workflows Qwen: the fastest open-weight reasoning model, now running in real time
Big thanks to the devs, startups, customers, and partners who made it electric. -
AI-Written Code and Developer Orchestration at Anthropic
By
–
Behind the scenes fireside with Anthropic and the press: “Over 70% new of pull requests are now AI written” “Developers are orchestrating Claude Code now”