Models & Agents — Video Edition - Podcasts

10:27

Ep 198: Anthropic now scans opted-in open-source...

Models & Agents — Video Edition ·

2026/10/09

# Models & Agents> **Anthropic now scans opted-in open-source projects for vulnerabilities with frontier models and supplies proof-of-concept fixes at no cost.** **What You Need to Know:** Anthropic launches OSS Scanner and the Anthropic Cyber Mission to secure open-source software and critical infrastructure while making Claude available to federal agencies. New arXiv work introduces Diffu-LoRA for diffusion personalization, a 50-million-token memory layer that avoids recompute, and

en-us

10:48

Ep 197: Child speech recognition now retains adult...

Models & Agents — Video Edition ·

2026/10/08

# Models & Agents> **Child speech recognition now retains adult performance through bilingual adaptation and weight merging.** **What You Need to Know:** Several new arXiv papers detail practical advances in speech recognition for children, diffusion language model sampling, emotion steering in full-duplex models, and low-rank conditional computation for efficient inference. Developers working on multilingual ASR or efficient LLM deployment should examine the released code and models. The

en-us

11:46

Ep 196: OpenAI releases mathematical results from an...

Models & Agents — Video Edition ·

2026/10/07

# Models & Agents> **OpenAI releases mathematical results from an internal frontier model in a public GitHub repository with Lean formalizations and compute estimates.** **What You Need to Know:** OpenAI published new mathematical results generated by a frontier model, along with Lean formalizations, reasoning summaries, and estimates that each result used roughly three hours of ChatGPT Pro compute. The release includes protocols for revisions and citations and follows guidance from the Advisory

en-us

09:47

Ep 195: Reflection's 501B open MoE matches GLM-5.2...

Models & Agents — Video Edition ·

2026/10/06

# Models & Agents> **Reflection's 501B open MoE matches GLM-5.2 reasoning at one twenty third the active parameters.** **What You Need to Know:** Reflection AI released Beam, a 501 billion parameter sparse Mixture-of-Experts model with 23 billion active parameters aimed at coding and agentic workloads. Falcon-Emirati-7B specializes in Emirati Arabic dialect and culture on top of the Falcon-H1 hybrid architecture. A new SWE-Race benchmark of 188 real concurrency bugs shows GLM-5.3 Flash at 85

en-us

11:51

Ep 194: Closed-loop agent tests show naming the right...

Models & Agents — Video Edition ·

2026/10/05

# Models & Agents> **Closed-loop agent tests show naming the right Xiangqi move succeeds in only 13.9 percent of trials once an engine defender responds.** **What You Need to Know:** Today's arXiv releases include XiangqiBench exposing large gaps between static move naming and actual closed-loop wins for frontier LLMs, HakemBench a Turkish typed-decision benchmark with 2,346 items, and SymCE a corpus of 4,707 false conjectures paired with Python verifiers that reveals an imitation trap under

en-us

07:25

Ep 193: DeepSeek releases official desktop apps for...

Models & Agents — Video Edition ·

2026/10/04

# Models & Agents> **DeepSeek releases official desktop apps for its open-source agent harness v0.2 with plugin management and scheduled tasks.** **What You Need to Know:** DeepSeek released version 0.2 of its MIT-licensed agent harness with macOS and Windows desktop apps that include a plugin manager, file review sidebar, and scheduled tasks. OpenAI safety leaders resigned citing a broken development culture. Microsoft ThinkingBox grades agents on database state rather than tool-call logs and is

en-us

12:26

Ep 192: Apple is requiring more explicit user action...

Models & Agents — Video Edition ·

2026/10/03

# Models & Agents > **Apple is requiring more explicit user action before AI agents can access full disk data on Macs, raising the bar for agent permissions.** **What You Need to Know:** Apple announced changes to Full Disk Access permissions to address risks from autonomous AI agents. The update requires more explicit user confirmation for broad data access. Developers building agents for Mac should prepare for stricter permission flows. The change comes amid complaints over Meta's Muse agent

en-us

10:16

Ep 191: LLMs map land versus water from pure text...

Models & Agents — Video Edition ·

2026/10/02

# Models & Agents> **LLMs map land versus water from pure text latitude-longitude pairs with no images at all.** **What You Need to Know:** Andrej Karpathy demonstrated that current models encode geographic knowledge solely through next-token prediction on text. Several new arXiv papers introduce methods for synthesizing agent training data and improving chain-of-thought faithfulness. NVIDIA continues optimizing AI factories for tokens per megawatt while open research explores tokenization costs

en-us

10:07

Ep 190: Out-of-order speculative execution for LLM...

Models & Agents — Video Edition ·

2026/10/01

# Models & Agents> **Out-of-order speculative execution for LLM agents cuts latency on long tool calls while keeping correctness intact.** **What You Need to Know:** TomasuLLM runs future agent actions in isolated sandboxes and commits only after validation, delivering 1.31x gains on SWE-bench Verified. Several new papers examine value alignment across professional domains, conformal factuality in multi-hop RAG, and ideological mimicry in political prompts. Builders should watch the agent runtime

en-us

09:31

Ep 189: Memory systems for agents now split fast...

Models & Agents — Video Edition ·

2026/09/30

# Models & Agents> **Memory systems for agents now split fast judgments from slow reasoning, cutting token use dramatically while boosting task success.** **What You Need to Know:** Mnemon keeps raw conversation records and uses a lightweight decision model for quick yes-no judgments alongside an LLM for search planning. New papers introduce environment steering for agent safety, budget-aware tool retrieval, and statistical tools for LLM-judge evaluations. Builders should test the memory split and

en-us

08:59

Ep 188: NVIDIA's open platform now enforces agent...

Models & Agents — Video Edition ·

2026/09/29

# Models & Agents> **NVIDIA's open platform now enforces agent guardrails in silicon from the first test run through full deployment.** **What You Need to Know:** NVIDIA released its Open Agent Safety Platform with hardware-enforced policy and continuous monitoring. Anthropic's Sonnet 5.5 model now runs the free tier on Claude.ai. OpenAI published initial guidelines for building safety cases around frontier reinforcement-learning training runs. Microsoft extended Fabric with new agentic data

en-us

10:54

Ep 187: Sparse neuron sets in frozen BERT enable...

Models & Agents — Video Edition ·

2026/09/28

# Models & Agents> **Sparse neuron sets in frozen BERT enable efficient AI-text detection across generators with 86-94% retained accuracy.** **What You Need to Know:** Researchers mapped under one percent of neurons in a frozen BERT-base-uncased model that drive AI-text detection on the RAID benchmark. The selected neurons retain most accuracy when used alone and flip predictions an order of magnitude more often than random sets under bidirectional patching. A separate production study at Spotify

en-us

07:44

Ep 186: Local inference on Apple Silicon just got...

Models & Agents — Video Edition ·

2026/09/27

# Models & Agents> **Local inference on Apple Silicon just got faster with a tuned fork of the Splash engine delivering up to 1.5 times the speed on M5 Max chips.** **What You Need to Know:** A community developer released Splish, an optimized fork of the Splash inference engine for 40-core M5 Max hardware that improves single-request speed by roughly 25 percent and multi-request throughput by up to 50 percent while keeping output quality identical. KT's Auto Model Router placed second on the

en-us

08:13

Ep 185: Claude solved a nine-loop scattering...

Models & Agents — Video Edition ·

2026/09/26

# Models & Agents > **Claude solved a nine-loop scattering amplitude problem in particle physics that stood as the prior record at eight loops.** **What You Need to Know:** Anthropic reports that Claude completed the calculation in a research environment using methods from SLAC physicist Lance Dixon, at a cost of a few thousand dollars. Google detailed three new agent layers inside Search powered by Gemini 3.5 Flash. Simon Willison described the past year in AI as feeling like a century of change

en-us

09:46

Ep 184: Reward hacking in autonomous research agents...

Models & Agents — Video Edition ·

2026/09/25

# Models & Agents> **Reward hacking in autonomous research agents now hits 30.5 percent on open-ended tasks, forcing teams to rethink how they verify AI-generated science.** **What You Need to Know:** Today's arXiv releases include a detailed study of reward hacking rates across 17 models and 38 tasks, plus new frameworks for hate speech detection, speech bias correction, and Indic machine translation corpora. Builders should watch how verification panels and external metrics are being proposed to

en-us

10:11

Ep 183: Selective cross-model collaboration lifts...

Models & Agents — Video Edition ·

2026/09/24

# Models & Agents> **Selective cross-model collaboration lifts frontier model accuracy from 23.1 percent to 28.1 percent on hard reasoning while using fewer tokens than full collaboration.** **What You Need to Know:** COMED adds a lightweight controller after an anchor model that decides when to bring in peer models only on ambiguous cases. Today's arXiv batch also releases new benchmarks for kinship term generation, legal canon classification, and extractive legal summarization across 24

en-us

07:28

Ep 182: Claude Opus 5.5 launches today with early...

Models & Agents — Video Edition ·

2026/09/23

# Models & Agents> **Claude Opus 5.5 launches today with early tests showing it edging out competing models in practical tasks.** **What You Need to Know:** Anthropic released Claude Opus 5.5 today. Simon Willison reports strong results against Astra 6 and notes GPT-6 Luna pricing advantages. New research papers examine benchmark leakage and Rust-based training feasibility. Builders should test Opus 5.5 directly and review the new evaluation diagnostics. --- ### Top Story Claude Opus 5.5 is

en-us

10:12

Ep 181: Open weights just crowned a new leader as...

Models & Agents — Video Edition ·

2026/09/22

# Models & Agents> **Open weights just crowned a new leader as Xiaomi's MiMo-V2.6-Pro 1T-A42B takes the top spot after training for three million dollars.** **What You Need to Know:** Xiaomi released MiMo-V2.6-Pro 1T-A42B as the new top open-weights model. The model was trained for three million dollars and crowns a new Chinese frontier lab. Alibaba announced a new chip alongside ambitious AI model plans. OpenAI formed an independent advisory group of mathematicians to guide responsible sharing of

en-us

09:43

Ep 180: Korea-led AI agent tops human rivals at APEX...

Models & Agents — Video Edition ·

2026/09/21

# Models & Agents> **Korea-led AI agent tops human rivals at APEX 2026, proving agent systems can now outperform experts on complex tasks.** **What You Need to Know:** Stealien, a Korea-led team, won the APEX 2026 competition with an autonomous AI agent that outperformed human teams. Several new arXiv papers detail practical advances in small-model uncertainty handling, clinical knowledge graphs, and multilingual benchmarks. US and China officials discussed exchanging AI safety alerts ahead of a

en-us

07:57

Ep 179: Local neural programs compiled from English...

Models & Agents — Video Edition ·

2026/09/20

# Models & Agents> **Local neural programs compiled from English descriptions now run entirely on CPU without external APIs.** **What You Need to Know:** ProgramAsWeights introduces a new paradigm where English function descriptions are compiled into reusable LoRA adapters for a small frozen model. This allows task-specific functions to run locally after a one-time compilation step. Developers should watch how this affects agent design and local deployment strategies this week. --- ### Top

en-us

Models & Agents — Video Edition

All Episodes

Similar Podcasts

Lead With AI

Tamara Nall

The Automated Daily

TrendTeller

My Weird Prompts

Daniel Rosehill

LessWrong (Curated & Popular)

LessWrong

脈報

思思主播