16th October - AI News Daily - Claude Haiku 4.5 Doubles Speed at One-Third Cost, Disrupts Agent Economics

Send us a text

🌍 INAI • The Open AI Hub

The Intelligence Atlas → the world’s most comprehensive, open hub of AI knowledge. 2 Million+ tools, models, agents, tutorials & daily news—free for all, updated every day.

https://github.com/inai-sandy/inAI-wiki

Top Highlights: Anthropic's Claude Haiku 4.5delivers faster, cheaper performance matching larger models on coding. Google DeepMind launches Veo 3.1 for AI video and teases Gemini 3.0 Pro. Microsoft unveils MAI-Image-1 for photorealistic images and an Agent Framework for DevOps. Walmart integrates instant checkout in ChatGPT while Salesforce+OpenAI bring CRM data to conversational workflows. Infrastructure expands with OpenAI+Oracle planning 450k GPUs, NVIDIA shipping DGX Sparks, and Meta starting a 1GW data center.

Tools: retrieve-dspyimproves retrieval pipelines; LlamaAgentssimplifies document extraction; GEPA+DSPyoffers auditable PII redaction; Ampprovides free agentic coding; Microsoft's Agent Framework SDKand Azure Local MCP Serverenable DevOps automation.

Models: Claude Haiku 4.5doubles speed at 1/3 cost; Veo 3.1adds audio and editing; MAI-Image-1targets photorealism; Samsung's TRMpacks reasoning in 7M parameters; Qwen3-Next-80Bruns efficiently on Apple hardware; GLM-4.6leads open coding benchmarks.

Research: Recursive Language Modelsenable unbounded context; thinking tokens researchreveals compute allocation patterns; Meta's ETDimproves reasoning; NVIDIA's PRM workenhances reward modeling; MALT datasetstudies reward hacking; EZSpecificityaccelerates drug discovery with 91% accuracy.

Industry: Salesforce+OpenAIintegrate Agentforce into ChatGPT; Walmart+OpenAIlaunch agentic commerce; OpenAI+Oracleplan 450k GPU deployment; NVIDIA and Metaexpand infrastructure; content authenticity efforts accelerate; OpenAIallows age-gated mature content.

Education: Tutorials cover Next.js voice transcription, Stanford's nanochat deep dive, LeRobotHF robotics guides, DSPy prompt optimization, and nanochat workflows.

Demos: ChatGPT ran Doom in-browser; Veo 3.1 stress-tested publicly; nanochat multimodal demoachieved sub-$10 training; Claude subagentsshowcased parallelized coding; HivergeAIset CIFAR-10 speed record.

Discussions: AGI timelinesface skepticism; Sora 2framed as participatory system; GPU export restrictionsmay limit innovation; verbalized samplingboosts creativity; methodology advancesinclude ColBERT tweaks and multimodal retrieval improvements.

Support the show