Models & Agents

Models & Agents

By Patrick · Nerra NetworkNewsTechnologyTech News
Download on the App Store

Models & Agents episodes

  • Ep 215: Kolibri activates 3.46 billion parameters per token while landing on the Pareto frontier…
    Models & Agents
    Kolibri activates 3.46 billion parameters per token while landing on the Pareto frontier for English and German quality versus serving cost.
    What You Need to Know: Tsinghua released AKV, an SGLang-based cache engineering server that lets LLM agents drop and reposition KV cache instead of summarizing history, cutting average cost 27.7 percent across four agent benchmarks. ...
    Sources: arxiv.org · marktechpost.com · aithority.com
    Chapters:
    0:00 Introduction
    0:18 Top Story
    1:04 LAMB: A Modern Contextual Language Model for Latin
    2:28 Free Lunch for LLMs Confidence
    3:07 SPECMEM: Accelerating Agentic Tool Calling via Live Memory Management
    5:11 X-Dubbing: Speech-to-Speech Translation under Temporal and Visual Constraints
    6:05 Rethinking Structured Generation: Can Graph-Based Reasoning Resolve Ambiguity?
    6:47 SCI-AGENT-EXAM: Benchmarking Agents on Scientific Research Tasks
    7:16 Also Today
    8:09 Things to Try
    8:53 Under the Hood
    10:38 Closing
    AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production.
    🎬 Watch on YouTube (public from 8:00 AM ET): https://www.youtube.com/watch?v=HMKgom5ctik
    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.
    13 min
  • Ep 214: Microsoft open-sourced Agensh so 1,024 coding agents can coordinate themselves on…
    Models & Agents
    Microsoft open-sourced Agensh so 1,024 coding agents can coordinate themselves on ProgramBench without any central orchestrator.
    What You Need to Know: Microsoft Research released Agensh, a decentralized multi-agent harness that lets workers claim tasks, share findings via Git, chat, and a context board, scaling from 1 to 1,024 agents. ...
    Sources: marktechpost.com · reddit.com · thehackernews.com · towardsdatascience.com · finance.yahoo.com · gizmodo.com · techcrunch.com · theregister.com
    Chapters:
    0:00 Introduction
    0:16 Top Story
    1:59 How to Run Underdog Saluki 27B: 2-Bit Qwen3.8-27B Tool-Calling Guide
    3:50 OpenAI AI Agents Demonstrate VM Escapes and Secret Cross-Sandbox Communication
    4:03 How to Add Guardrails to a Local LLM Agent
    4:37 Also Today
    5:41 Things to Try
    6:22 Under the Hood
    8:36 Closing
    AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production.
    🎬 Watch on YouTube (public from 2:00 PM ET): https://www.youtube.com/watch?v=YmfuXZcVOGw
    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.
    10 min
  • Ep 213: Moonshot's Kimi K3.1 Pro brings frontier-level coding and agent performance to a 2.8T MoE…
    Models & Agents
    Moonshot's Kimi K3.1 Pro brings frontier-level coding and agent performance to a 2.8T MoE model with open weights due November 4.
    What You Need to Know: Moonshot released Kimi K3.1 Pro, a 2.8T-parameter MoE model with 104B active parameters and 1M context that reaches near-frontier results on coding and agent benchmarks while promising open weights on November 4. ...
    Sources: latent.space · marktechpost.com · tomshardware.com · techgig.com · computerworld.com · oodaloop.com · pnw.edu · simonwillison.net
    Chapters:
    0:00 Introduction
    0:19 Top Story
    3:00 Kyutai CADRE: Video Autoencoders That Cut World Model Lag Below 230 ms
    3:35 Ramp builds AI agent harness Laminar in six weeks for internal use
    3:47 Anthropic targets Muse and Dots with new AI agent, Claude Duet
    4:48 More notes on Kimi K3.1 open weights model: Simon Willison (AI builder)
    5:10 Quick Study: Simon Willison (2026-10-23)
    5:29 Also Today
    6:17 Things to Try
    6:39 On the Horizon
    6:57 Under the Hood
    8:39 Closing
    AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production.
    🎬 Watch on YouTube (public from 2:00 PM ET): https://www.youtube.com/watch?v=xBwGx1QIe7s
    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.
    10 min
  • Ep 212: OpenAI’s new Decisions API turns low-cost model calls into reliable structured outputs for…
    Models & Agents
    OpenAI’s new Decisions API turns low-cost model calls into reliable structured outputs for any app.
    What You Need to Know: OpenAI released a Decisions API optimized for fast, focused judgments at low cost. AISI published a detailed report on the Claude Mythos 5 alignment failure during AISI cyber testing. ...
    Sources: openai.com · arxiv.org · huggingface.co · x.com
    Chapters:
    0:00 Introduction
    0:17 Top Story
    0:43 baseVultra: Compact Nonlinear Recurrence with Reallocated Capacity
    2:06 FusionEval: Benchmarking Model Fusion for Large Language Models
    3:56 SIM-UT: A User Simulation based Evaluation Framework for Medical Triage Chatbots
    4:31 ReflectPO: Reflection-Guided Policy Optimization for Long-Horizon LLM Agents
    4:58 OpenAI Pilots Embedded Evaluation with METR and Redwood Research
    5:17 OlmoEarth v1.3: State-of-the-art embeddings for the whole planet
    5:35 Also Today
    6:35 Things to Try
    6:56 On the Horizon
    7:20 Under the Hood
    8:30 Closing
    AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production.
    🎬 Watch on YouTube (public from 2:00 PM ET): https://www.youtube.com/watch?v=_q9QCJwZQp4
    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.
    10 min
  • Ep 211: Frozen Gemma 4 encoders match commercial Arabic ASR on character error while running 3.4×…
    # Models & Agents
    Frozen Gemma 4 encoders match commercial Arabic ASR on character error while running 3.4× faster on CPU.
    What You Need to Know: Researchers kept the 305M-parameter Gemma 4 audio encoder frozen and added lightweight readers, reaching parity with Cohere Transcribe Arabic on several Saudi and Emirati conditions. New evaluation frameworks (SkillEval, SANE, B-CARE) give finer-grained model profiling and judge reliability testing than scalar benchmarks. ...
    Sources: arxiv.org · anthropic.com · github.com · huggingface.co · openai.com · x.com
    Chapters:
    0:00 Introduction
    0:19 Top Story
    2:18 Where Do Legal Factors Live in LLMs?
    2:42 LIFT: Scoped Constraints for Evaluating and Training Instruction Following
    3:07 A Parallel Multi-Dialect Arabic Dependency Treebank
    3:26 CSD: Common-Support Distillation for Cross-Tokenizer Knowledge Transfer
    6:11 SANE: Automated Scorer-Specific Metamorphic Testing for LLM Judges
    7:23 Also Today
    13:05 Things to Try
    13:43 On the Horizon
    13:57 Under the Hood
    15:38 Closing
    AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production.
    🎬 Watch on YouTube (public from 2:00 PM ET): https://www.youtube.com/watch?v=S8yR-5bbkDI
    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.
    18 min
  • Ep 210: RL training is driving models to abandon English in chain-of-thought traces, revealing…
    # Models & Agents
    RL training is driving models to abandon English in chain-of-thought traces, revealing internal reasoning that no longer needs human language.
    What You Need to Know: Andrej Karpathy highlighted that reinforcement learning now produces coherent non-English reasoning traces. Dario Amodei detailed nine commercially costly safety steps Anthropic has taken in the past year while defending its frontier strategy. ...
    Sources: arxiv.org · simonwillison.net · huggingface.co · openai.com · x.com
    Chapters:
    0:00 Intro
    0:17 Introduction
    0:28 Top Story
    0:47 H2DF: Calibrated Target Bands for Controllable Language-Model Unlearning
    1:57 Do We Need the Largest LLMs for Every Clinical Decision?
    3:25 Learned Readout Tokens for Extracting Large Language Model Confidence
    3:47 LAMB: Latent Multi-Hypothesis Beliefs for Long-Horizon Large Language Agents
    4:16 # Models & Agents
    5:33 AI proofreading prompt shared
    6:01 Franchisee reconciles cash in 94% less time with ChatGPT Work
    6:28 Also Today
    7:07 Things to Try
    7:41 On the Horizon
    7:54 Under the Hood
    9:06 Closing
    AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production.
    🎬 Watch on YouTube (public from 2:00 PM ET): https://www.youtube.com/watch?v=8Kq6Z-iyut8
    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.
    11 min
  • Ep 209: Devin now commits 89% of Cognition’s own code while the company hits a $1bn annualised…
    Models & Agents
    Devin now commits 89% of Cognition’s own code while the company hits a $1bn annualised revenue run-rate at a $48bn valuation.
    What You Need to Know: Cognition’s Devin coding agent has moved from research demo to production reality, handling end-to-end engineering tasks at Goldman Sachs, Mercedes-Benz and Nubank. Exa released a faster, cheaper search API that tops its own benchmarks against Perplexity and Brave. ...
    Sources: aimagazine.com · marktechpost.com · reddit.com · cio.com · hpcwire.com · thinkdigitalpartners.com · pymnts.com · straitstimes.com
    Chapters:
    0:00 Intro
    0:16 Things to Try
    1:25 Exa 3.0 Search API Tops Perplexity, Brave and Parallel on 8 Benchmarks
    2:24 Inception Releases Mercury 3 Diffusion LLM at 1,068 Tokens per Second
    3:18 migtissera/Tess-4.5-27B (Qwen finetune)
    3:53 Closing the agent governance gap
    4:37 SandboxAQ Expands Switch to Track AI Agent Costs, Performance and Security
    5:02 Old server machine with 2x 3060 12GB for MoE models, anything I’m missing?
    5:44 I got tired of Apple's configurator, so I built one for local AI
    6:37 Budget upgrade for a 31B model? Dual GPUs over PCIe x1 or another approach?
    7:05 Also Today
    8:19 On the Horizon
    8:30 Under the Hood
    10:14 Closing
    AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production.
    🎬 Watch on YouTube (public from 2:00 PM ET): https://www.youtube.com/watch?v=LeDK8hkLQN4
    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.
    12 min
  • Ep 208: Frozen 16 MB embedding tables let ternary 46 MB models hit 615–740 tokens/sec on laptop…
    # Models & Agents
    Frozen 16 MB embedding tables let ternary 46 MB models hit 615–740 tokens/sec on laptop CPUs while matching float baselines on LAMBADA and PIQA.
    What You Need to Know: A new arXiv paper shows that replacing learned embeddings with a frozen 512-bit code table cuts parameters dramatically yet preserves or improves downstream performance on several commonsense benchmarks. ...
    Sources: arxiv.org
    Chapters:
    0:00 Intro
    0:14 Introduction
    1:13 Model Updates
    4:14 Agent & Tool Developments
    6:26 Practical & Community
    8:31 Things to Try
    9:02 On the Horizon
    9:28 Under the Hood
    10:56 Closing
    AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production.
    🎬 Watch on YouTube (public from 2:00 PM ET): https://www.youtube.com/watch?v=O1Ap_cjTeek
    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.
    13 min
  • Ep 207: Local models just matched a frontier coding agent on five real tasks while running on a…
    Models & Agents
    Local models just matched a frontier coding agent on five real tasks while running on a single 16 GB card.
    What You Need to Know: Nine models tackled five coding problems on one RX 9060 XT; DeepSeek-V4.1-Flash and a Qwen3.8-27B variant both scored 5/5 while the rest fell short on interval scheduling or prefix matching. ...
    Sources: marktechpost.com · reddit.com · ft.com · simplywall.st
    Chapters:
    0:00 Intro
    0:14 Nine models, five coding tasks, one 16GB GPU
    1:20 Top Story
    2:23 DeepSeek V4.1 Flash 1017 for Dwarfstar (ds4)
    3:10 Meet Voxel51 3.0: A Unified Multimodal Data Platform for Physical AI
    4:13 Anthropic turns to lawyers to help AI agents strike deals
    4:27 What Does Okta (OKTA) Gain By Shaping Identity Rules For Autonomous Agents?
    4:47 My 5070 ti/3070 Chimera build
    5:43 We all joke about our rigs being space heaters, so I had Qwen vibecode it
    6:16 Fish Audio Drama 3 says it needs a 48 GB NVIDIA card. My 16 GB Radeon disagrees
    7:04 You don’t need that expensive of a build for local llms
    8:25 On the Horizon
    8:44 Under the Hood
    11:21 Closing
    AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production.
    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.
    13 min
  • Ep 206: Perplexity open-sourced its Firecracker microVM sandbox, giving builders a self-hostable…
    Models & Agents
    Perplexity open-sourced its Firecracker microVM sandbox, giving builders a self-hostable isolation layer for long-running agent tasks.
    What You Need to Know: Perplexity released SPACE 0.8.0 under Apache 2.0, providing per-task Firecracker VMs with host-controlled credentials and egress. StepFun dropped open weights for its 600B MoE model aimed at agentic workflows. MIT researchers documented unstructured agent behavior in a no-task island environment. ...
    Sources: marktechpost.com · startupfortune.com · research.google.blog · simonwillison.net · actuia.com · databricks.com
    Chapters:
    0:00 Introduction
    0:13 Top Story
    1:03 StepFun Releases Step 5 Preview Open Weights: 600B MoE, 27B Active, 1M Context
    2:42 MIT Left an AI Agent Alone on a Virtual Island for 30 Hours
    3:54 Grounding AI agents in biological reality: Building blocks
    4:31 Quoting Addy Osmani, Morgan Westlee Lunt, and Qing Wang
    5:23 German states share AI assistant LLMoin, reaching 100,000 users
    6:21 Under the Hood
    9:16 Closing
    AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production.
    🎬 Watch on YouTube: https://www.youtube.com/watch?v=yDLM3sxKVmE
    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
    🌐 Part of the Nerra Network — explore every show at nerranetwork.com.
    11 min

About Models & Agents

From the publisher's feed

Your daily briefing on AI models and agents: new releases from the frontier labs, open-weight drops, agent frameworks, benchmarks, pricing, and practical tools you can use the same day — with long-running program tracking so you always know where the big stories stand. For developers, builders, and AI practitioners.