Ep 225: Together Compiler lets agents re-optimize inference on the fly and delivers up to 2.6…

Models & Agents Vercel’s UI SDK supplies agents with 50 behaviors for building accessible components across five frameworks. What You Need to Know: Vercel Labs released an Apache-2.0 beta library of 50+ UI behaviors that agents compose into components for React, Vue, Svelte, Solid, and Web Components. Agents using the bridge cut tool calls by 69% and task time by 61% on a settings task in Vercel’s tests. ... Sources: marktechpost.com · reddit.com · digitaltrends.com · ukrmedia.news · anthropic.com · finance.biggo.com · towardsdatascience.com Chapters: 0:00 Introduction 0:18 Top Story 1:28 I have trained a model to predict my blood sugar [P] 2:27 Bolt PCHIE: a deterministic, tamper-evident safety harness for AI agents [P] 2:48 Perplexity Teammates bring autonomous AI agents to the Mac app 3:17 OpenAI has launched ChatGPT Wallet for AI-powered purchases 3:58 10 Recommendations for Building Strong Safety Evaluations 4:38 Google Releases TPU Sync: Open-Source KV Cache Transfer for TPUs 5:48 Things to Try 6:23 On the Horizon 6:45 Scaling to 1T MoE on TorchTitan [P] 8:21 Closing AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production. 🎬 Watch on YouTube (public from 8:15 AM ET): https://www.youtube.com/watch?v=ojfYfADImj0 If this episode was useful, a rating or a short review on Apple Podcasts or Spotify is how the next listener finds the show — and following it in your app means the next episode is there when you are. 📝 Full show notes, transcript & sources: read the episode page 🎧 Hear this show with the others you follow in one private morning episode, read to you by name. Try My Nerra free for 7 days, then $4.99 a month. 📬 The free Nerra Daily email, every morning at 7 AM Eastern: sign up 🌐 Part of the Nerra Network — explore every show at nerranetwork.com.

Models & Agents Gemini 4 Hydrogen ships at $1.50/$7.50 per million tokens and powers the free tier plus Google Search. What You Need to Know: Google released Gemini 4 Hydrogen, the lowest-cost member of the Gemini 4 family, positioned for everyday coding and agent execution while Carbon and Argon handle deeper reasoning. ... Sources: blog.google · simonwillison.net · reddit.com · ipo.anthropic.com · marktechpost.com · designnews.com · hpcwire.com · pasqualepillitteri.it Chapters: 0:00 Introduction 0:19 Top Story 1:21 llama.cpp bumps version to 1.0.0 1:53 Ernie 5.5 Flash Preview launches at 280B total parameters with 14B active 2:15 Anthropic files for initial public offering 2:48 A100 prices fall 20-30% since earlier collapse reports 3:09 CopilotKit Releases OpenBot 1.0 3:28 Limitless Labs CAM Agent arrives for engineers 3:40 Strata EXL3 AMD fork gains 10-20% decode speed on HIP kernels 4:53 RAG retrieval lost 40 points when distracting passages grew to 2.68M 5:15 LLMON: An LLM-Native Markup Language now in Granite 4.1 3B and 8B 6:13 Also Today 7:04 Things to Try 7:59 On the Horizon 8:25 Closing AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production. 🎬 Watch on YouTube (public from 8:15 AM ET): https://www.youtube.com/watch?v=NwaUAA322ug If this episode was useful, a rating or a short review on Apple Podcasts or Spotify is how the next listener finds the show — and following it in your app means the next episode is there when you are. 📝 Full show notes, transcript & sources: read the episode page 🎧 Hear this show with the others you follow in one private morning episode, read to you by name. Try My Nerra free for 7 days, then $4.99 a month. 📬 The free Nerra Daily email, every morning at 7 AM Eastern: sign up 🌐 Part of the Nerra Network — explore every show at nerranetwork.com.

Models & Agents Google's Gemini Decision 1 scores up to 1,000 multimodal questions in one 150-250 ms pass at $0.10 per million input tokens with zero output cost. What You Need to Know: Google launched the Gemini Decisions API powered by Gemini Decision 1, a specialized model that returns calibrated probabilities for yes/no, choice, and rubric questions over text, images, audio, video, and PDFs without generating tokens. ... Sources: simonwillison.net · alphagenomedocs.com · marktechpost.com · smbtech.au · aws.amazon.com · scworld.com · pulse2.com · ynetnews.com Chapters: 0:00 Introduction 0:18 Top Story 1:10 Announcing AlphaGenome 2 DNA sequence model 1:42 Google DeepMind Releases AlphaGenome 2 for Cell-Type-Specific Variant Effects 2:58 Google releases open-source agent substrate for secure AI sandboxing 3:13 aws-emulator: Simon Willison (AI builder) 3:39 WhiteMatter: $20 Million Series A To Expand Enterprise AI Agent Governance 4:07 Plurai raises $13M to catch AI agent failures before customers do 4:23 How to Use Apple's fm CLI to Run On-Device LLMs From Python 4:39 Also Today 5:23 Things to Try 5:57 On the Horizon 6:09 Under the Hood 7:36 Closing AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production. 🎬 Watch on YouTube (public from 8:15 AM ET): https://www.youtube.com/watch?v=oZf4eyDQQnk If this episode was useful, a rating or a short review on Apple Podcasts or Spotify is how the next listener finds the show — and following it in your app means the next episode is there when you are. 📝 Full show notes, transcript & sources: read the episode page 🎧 Hear this show with the others you follow in one private morning episode, read to you by name. Try My Nerra free for 7 days, then $4.99 a month. 📬 The free Nerra Daily email, every morning at 7 AM Eastern: sign up 🌐 Part of the Nerra Network — explore every show at nerranetwork.com.

Models & Agents A new 0.8B on-device model doubles agentic benchmark scores while keeping all data private. What You Need to Know: Tether AI released BotPsy, an 0.8B model fine-tuned from Qwen3.5-0.8B for function calling and agentic tasks across web, travel, finance, and device control. It roughly doubles performance on BFCL v4, ACEBench, and τ-bench compared with its base while running entirely locally. ... Sources: huggingface.co · arxiv.org · openai.com · blog.google · anthropic.com Chapters: 0:00 Introduction 0:14 Top Story 5:47 Businesses in ChatGPT now in beta with Business Agents and engagement insights 6:37 OpenAI Academy expands free ChatGPT workshops to caregivers in five U.S. cities 6:54 Anthropic begins its IPO roadshow with a filed S-1 registration statement 7:10 Also Today 9:35 Things to Try 10:05 On the Horizon 10:26 Closing AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production. 🎬 Watch on YouTube (public from 8:15 AM ET): https://www.youtube.com/watch?v=f_UeZW44xQs If this episode was useful, a rating or a short review on Apple Podcasts or Spotify is how the next listener finds the show — and following it in your app means the next episode is there when you are. 📝 Full show notes, transcript & sources: read the episode page 🎧 Hear this show with the others you follow in one private morning episode, read to you by name. Try My Nerra free for 7 days, then $4.99 a month. 📬 The free Nerra Daily email, every morning at 7 AM Eastern: sign up 🌐 Part of the Nerra Network — explore every show at nerranetwork.com.

Models & Agents Anthropic cut Fable 5.2 prices 20 percent while raising its FrontierCode 2.0 score from 32.4 to 49.1 percent at xhigh effort. What You Need to Know: Anthropic released Claude Fable 5.2 and Mythos 5.2 with 30 percent faster output, $8 per million input and $40 per million output tokens, and a new prompting pattern that tells the model the codebase belongs to it so it merges rather than duplicates code. ... Sources: computerworld.com · marktechpost.com · prnewswire.com · scworld.com · pymnts.com · itpro.com · simonwillison.net · reddit.com Chapters: 0:00 Introduction 0:25 Top Story 1:10 Meta Releases Muse Ember, a 1M-Context Model for High-Volume Agent Work 2:15 Google Unveils TPU 8i Max and TPU 8tx: A Dual-Chip TPU and a Confidential TPU 2:44 OpenAI Introduces Control Room System for Model Workload Tracking and Safety 3:24 Amazon and Walmart Bring AI Agents Closer to the Purchase 3:40 Capgemini to lead EU consortium for AI safety and cyber risk assessment 4:35 Quoting Prompting Claude Fable 5.2: Simon Willison (AI builder) 4:53 Yonkers Rises: Simon Willison (AI builder) 5:10 Also Today 5:53 Things to Try 6:32 On the Horizon 8:06 Closing AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production. 🎬 Watch on YouTube (public from 8:15 AM ET): https://www.youtube.com/watch?v=L2Fij2bT7Pk If this episode was useful, a rating or a short review on Apple Podcasts or Spotify is how the next listener finds the show — and following it in your app means the next episode is there when you are. 📝 Full show notes, transcript & sources: read the episode page 📬 Get the sources and notes by email, free: join the newsletter 🌐 Part of the Nerra Network — explore every show at nerranetwork.com.

Models & Agents LG just opened free API access to its 750B K-EXAONE 2.5 model for agentic workloads. What You Need to Know: LG AI Research released free hosted access to K-EXAONE 2.5, a 750B-parameter Mixture-of-Experts model focused on Korean and agent tasks. It improves on the prior version across software development, web browsing, SaaS automation, and agent problem-solving benchmarks. ... Sources: mk.co.kr · pymnts.com · arxiv.org · simonwillison.net · statnews.com · cio.com · x.com Chapters: 0:00 Introduction 0:19 Top Story 1:18 Compression of Emotion in Language Model Representations 4:03 Also Today 12:00 Things to Try 12:24 On the Horizon 12:43 Closing AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production. 🎬 Watch on YouTube (public from 8:15 AM ET): https://www.youtube.com/watch?v=_oeIrskUR4E If this episode was useful, a rating or a short review on Apple Podcasts or Spotify is how the next listener finds the show — and following it in your app means the next episode is there when you are. 📝 Full show notes, transcript & sources: read the episode page 📬 Get the sources and notes by email, free: join the newsletter 🌐 Part of the Nerra Network — explore every show at nerranetwork.com.

Models & Agents Ten open MoE models expose routing choices that swing sparsity from 4x to 128x and change which hardware setups actually fit. What You Need to Know: A verified table from model configs shows DeepSeek-V4-Pro using 384 experts with sqrtsoftplus gating and hash routing in three layers, while Llama 4 Maverick keeps its gating method undisclosed. ... Sources: marktechpost.com · reddit.com · towardsdatascience.com · the-decoder.com · nytimes.com · thechenabtimes.com · karpathy.ai · casrai.org Chapters: 0:00 Introduction 0:19 Top Story 1:30 GRFNet, PyTorch model for multichannel time-series classification [P] 2:12 Timestral, t0, Tabby, and Morpheus-2.0 Compared on GIFT-Eval 2:36 Locally trained VLA model with minimal attention based vision encoder. [p] 3:46 Law Firms Have Malpractice Insurance. Should Their A.I. Agents? 4:03 Agent & Tool Developments 4:47 TristanTube: YouTube viewer from Tristan's top channels 5:20 Also Today 6:32 Things to Try 7:03 On the Horizon 8:05 Closing AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production. 🎬 Watch on YouTube (public from 8:15 AM ET): https://www.youtube.com/watch?v=sSe-P7Qr27E If this episode was useful, a rating or a short review on Apple Podcasts or Spotify is how the next listener finds the show — and following it in your app means the next episode is there when you are. 📝 Full show notes, transcript & sources: read the episode page 📬 Get the sources and notes by email, free: join the newsletter 🌐 Part of the Nerra Network — explore every show at nerranetwork.com.

Models & Agents Ling-3.2-flash ships a 560B/25B multimodal MoE with 1M context and OpenAI-compatible APIs for immediate use. What You Need to Know: Ant Ling released Ling-3.2-flash, a native multimodal mixture-of-experts model supporting text, image, and video with a 1M-token window. Buoyant open-sourced Kinference, a Kubernetes control plane that adds KV-cache-aware routing and policy enforcement on top of vLLM. ... Sources: reddit.com · nature.com · marktechpost.com · anthropic.com · huggingface.co · towardsdatascience.com · bioengineer.org · jpost.com Chapters: 0:00 Introduction 0:17 Top Story 0:59 Model Updates 2:58 OpenAI introduces BioBasicsBench for AI in wet-lab science 3:46 Build Your First Multi-Tool AI Agent with Ollama and Python 4:03 Give Your Python AI Agent Persistent Memory with One JSON File 4:21 Also Today 5:29 Things to Try 6:04 On the Horizon 7:52 Closing AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production. 🎬 Watch on YouTube (public from 8:15 AM ET): https://www.youtube.com/watch?v=bA_FLgX90Wg If this episode was useful, a rating or a short review on Apple Podcasts or Spotify is how the next listener finds the show — and following it in your app means the next episode is there when you are. 📝 Full show notes, transcript & sources: read the episode page 📬 Get the sources and notes by email, free: join the newsletter 🌐 Part of the Nerra Network — explore every show at nerranetwork.com.

# Models & Agents Gemini Omni 1.2 Flash handles longer video clips on edge devices at lower cost than prior Gemini versions. What You Need to Know: Google released Gemini Omni 1.2 Flash with improved handling of extended multimodal inputs. Kakao launched a decision model 43 times faster than GPT-6 Luna. ... Sources: blog.google · mbiz.heraldcorp.com · arxiv.org · itbrief.co.nz · enterprisetimes.co.uk · homeland411.com · thenational.co.za · ynetnews.com Chapters: 0:00 Introduction 0:16 Top Story 0:38 Kakao releases AI decision model 43 times faster than GPT-6 Luna 2:14 Web Hole attack targets AI agents via abandoned sites 2:29 Atsign EdgeAI Launches Secure Agent Connectivity 2:44 Microsoft Warns Agentic AI Weaponized to Steal Cloud Secrets in 5 Minutes 4:00 Dataset Collection and Fine-Tuning of NLLB for German--English Translation 4:21 CORE: Coarse-to-Fine Search with Context Management for Deep Research 4:38 Also Today 8:29 Things to Try 9:13 On the Horizon 9:31 Tomorrow Teaser 9:41 Closing AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production. 🎬 Watch on YouTube (public from 8:15 AM ET): https://www.youtube.com/watch?v=axdYeFfg-Ms If this episode was useful, a rating or a short review on Apple Podcasts or Spotify is how the next listener finds the show — and following it in your app means the next episode is there when you are. 📝 Full show notes, transcript & sources: read the episode page 📬 Get the sources and notes by email, free: join the newsletter 🌐 Part of the Nerra Network — explore every show at nerranetwork.com.

Models & Agents Together Compiler lets agents re-optimize inference on the fly and delivers up to 2.6 times higher throughput. What You Need to Know: OpenAI released GPT-Image-2.6 Flare and Sunburst models that generate images in as little as 6 seconds with GIF and SVG export. Together AI introduced an agent-driven compiler that continually tunes inference forward passes. ... Sources: marktechpost.com · arxiv.org · cio.com · www-cdn.anthropic.com · x.com Chapters: 0:00 Introduction 0:21 Top Story 2:38 Prospective Memory for Proactive LLM Agents 3:01 Cognitive Bias Signatures in Search-Equipped LLM Agents 3:59 RAGCompass: Generalizable Retriever Routing via Reranking 4:22 Editing Large Language Models with Geometry-Aware Regularization 4:47 Also Today 13:47 Things to Try 14:31 On the Horizon 14:51 Under the Hood 16:16 Closing AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production. 🎬 Watch on YouTube (public from 8:15 AM ET): https://www.youtube.com/watch?v=YrNfCMAiRog If this episode was useful, a rating or a short review on Apple Podcasts or Spotify is how the next listener finds the show — and following it in your app means the next episode is there when you are. 📝 Full show notes, transcript & sources: read the episode page 📬 Get the sources and notes by email, free: join the newsletter 🌐 Part of the Nerra Network — explore every show at nerranetwork.com.