🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

It's Hard Fork Friday! This week on the show, we take a field trip to Google I/O, and sit down with Demis Hassabis (our first Nobel laureate) for a fascinating conversation about preparing for AGI, AlphaEvolve and what comes next in AI.

"When we recognize code as elegant, other code becomes sloppy. When we praise efficiency, the notion of waste is born… Therefore The Vibe Coder builds without laboring" I'm sorry but, pretty as it as, this makes no sense at all.

I was reading about llm-d, a new open-source, K8s-native framework for high-performance distributed LLM inference from Google, IBM, & Red Hat! Key Features: > Advanced Caching & Routing: Smart scheduling (via vLLM Optimized Inference Scheduler) to KV-cached replicas. >

Veo 3 is rolling out to Pro tier! the shipping continues in Gemini web

Shocker! Claude 4 system prompt was leaked, and it's a goldmine! The Claude system prompt incorporates several identifiable agentic AI patterns as described in "A Pattern Language For Agentic AI." Here's an analysis of the key patterns used: Run-Loop Prompting: Claude

🔥🔥Best Way to Construct Veo 3 Prompts (The Dave Clark Way!) Think of constructing a Veo prompt like writing a very short, hyper-specific film script that only describes what the camera sees and feels, assuming your AI "cinematographer" knows nothing implicitly. Here's a

The amazing team at Google strikes again! While I did not work on Gemini, I had lots of fun collaborating closely with Googlers and explored one flavor of “parallel thinking” we call PASTA, check this out: arxiv.org/abs/2502.11517

I put together an annotated version of the new Claude 4 system prompt, covering both the prompt Anthropic published and the missing, leaked sections (thanks, Pliny the Liberator 🐉󠅫󠄼󠄿󠅆󠄵󠄐󠅀󠄼󠄹󠄾󠅉󠅭) that describe its various tools It's basically the secret missing manual for Claude 4, it's fascinating!

Google Veo3 T2V Prompting Guide. For consistent iterations. TIPS: + Start with a structure + Create a base prompt + Iterate small elements BASE PROMPT STRUCTURE: [PERSPECTIVE], [SHOT STYLE + DETAILS], [SUBJECT DETAILS + ENVIRONMENT], [SCENE DETAILS], [SOUND DESCRIPTION],

BIG from Google DeepMind. «DataRater … estimates the value of training on any particular data point. This is done by meta-learning using ‘meta-gradients’, with the objective of improving training efficiency on held out data». Of course they report substantial gains.

Announcing Mistral AI Agents API, a game-changer in building agents and putting AI into everyone’s hands! 💻 Code execution: agents can run Python code in a secure environment. 🎨 Image generation: create visual aids, custom graphics, and artistic images. 📚 Document library:

Gemini 2.5 Flash Preview now supports native audio output via the Live API for seamless, natural spoken interactions and greater voice control. A new experimental thinking version of this audio model supports reasoning capabilities for more complex tasks. ai.google.dev/gemini-api/doc…

"RL from a single example works" "RL with random rewards works" "Base model pass@256 can match RL model pass@1" "RL updates a small % of params" Recent papers all point in the same direction: RL is mostly just eliciting latent behavior already learned in pretraining, not

Super excited to announce Mistral AI Codestral Embed, our first embedding model specialized for code. It performs especially well for retrieval use cases on real-world code data.

I think "iteration" is a very important area when using generative AI. However, I also think that surprisingly many people do not perform this iteration. "Iteration until the desired result is obtained" is good, but even if the desired result is obtained, it may be good to try

Vibe Coding (generated by Veo 3)

speeding up neural net inference in C

🚀HUGE HUGE NEWS: Modal just committed $300,000+ in compute credits for the Agents & MCP Hackathon 2025! 😱All participants get $250 in GPU credits -- build & deploy serious AI apps at the hackathon and win BIG. Join now!! With Anthropic, Mistral AI, Hugging Face, &

The methods we used to trace the thoughts of Claude are now open to the public! Today, we are releasing a library which lets anyone generate graphs which show the internal reasoning steps a model used to arrive at an answer.

We’ve just published a new example of using dstack with Hugging Face TRL to train across multiple nodes! It lets you quickly run training on any cloud GPU or on-prem cluster—no K8s or Slurm needed. Check it out: dstack.ai/examples/distr…