Models & Agents

Models & Agents

What You Need to Know: AbridgeKV trains pause tokens to compress completed reasoning blocks into compact KV entries, matching full-cache accuracy on 446 competition problems while cutting memory and latency. Several new papers introduce reusable control layers, topology generators, and medical report analyzers that separate domain logic from execution. ...

Sources: arxiv.org

Chapters:

0:00 Intro

0:16 Top Story

1:50 Wolof-Adapted Industry-Ready Open-Weight Language Models

2:23 Verdict: Encode Once, Decide Many

3:01 APRiL: An Open-ended Benchmark for Chinese Social Awareness

3:51 DeepUnfoldingAgent: A ClassicalQuantum Adaptive Control API

4:45 MeddeX: A Retrieval-Augmented LLaMA-3 System with Gemini Vision

5:25 ReviewPilot: A Multi-Agent System for Screening and Data Extraction

6:37 SHIP: Structured and Hierarchical Insight with Privacy

8:41 PRISM: PRIvileged Sibling-guided Multi-Agent Topology Design

9:19 Under the Hood

11:04 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

🌐 Part of the Nerra Network — explore every show at nerranetwork.com.

Models & Agents

What You Need to Know: A 9B model fine-tuned with FlowSpec lifts documented compliance checks from 67-71% to 93-100% on four held-out debt-collection procedures. ...

Sources: arxiv.org

Chapters:

0:00 Intro

0:15 Top Story

1:26 LLM-jp-4 Technical Report

2:34 Copy-Paste Language Models

3:14 Qwenfant: Train Big Character-level BabyLMs on Concise Parallel Text

4:03 HERO: Harnessing Evidence for Test-time Refinement and Prompt Optimization

5:33 NEPSEE: A Double-Annotated Benchmark for Nepali Sentiment Analysis

7:05 Whose Data, Whose Benchmark? A User-Centric Framework

8:06 Titan-Llama: Retrofitting Pre-trained LLMs with External Neural Memory Modules

8:44 Under the Hood

10:41 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=5rGSZr41F0Q

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.

Models & Agents

What You Need to Know: Nvidia pledged $250 million in accelerated computing resources over five years to the UK for agentic safety testing and safeguards research. Workday launched Sana as the default interface that orchestrates secure agents for tasks like hiring and payroll. ...

Sources: ca.finance.yahoo.com · marktechpost.com · huggingface.co · itpro.com · itnews.asia · reddit.com

Chapters:

0:00 Intro

0:10 Introduction

1:08 Google Releases Gemini 4 Argon API With 1M Token Output

1:48 Run fast, private agents on-device with LFM2.5-24B-A2B

3:17 Workday puts agentic AI front and center

4:21 Updated list of official MCP servers (300+)

4:50 Train LoRA directly on Quantized GGUFs

5:43 Moving a running model's layers between GPUs mid-reply (8

6:32 Discover AI: A Learning Guide - 51 hands-on labs

8:18 Under the Hood

9:53 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=3MrC4oGk1dA

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.

Models & Agents

What You Need to Know: New arXiv work introduces POrchestra for proactive agent coordination and T²Mem for test-time memory building. Meta released models to power its Muse agent. Prompt optimizers and LLM-judge rubrics both show measurable side effects on requirements they were never scored on. ...

Sources: techgig.com · arxiv.org · community.nasscom.in · pymnts.com · latent.space · simonwillison.net · github.blog · businesstimes.com.sg

Chapters:

0:00 Intro

0:18 Why Multi-Agent AI Succeeds Where Single-Agent Systems Hit

1:18 Do Prompt Optimizers Break Requirements They Are Not Scored

1:49 MoPrO-SD: Modular Prompt Optimization for Stance Detection

2:22 Host-Aware Web Crawling for LLM Pretraining Data

3:26 Personalization Without Stereotyping: Do Speech LLMs Know

4:11 Mashreq’s AI Agents Cut Integration Time Up to 70%

4:49 Notes on Python 3.15, uv, and Python 3.11 features

5:59 Towards the Roof of the World: NLP Progress

7:22 Under the Hood

9:12 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=dg94PWrwWx4

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.

Models & Agents

What You Need to Know: SWE-Duel pits coding agents against each other in mirrored Red-Blue turns on real repositories, generating fresh tasks and oracles at evaluation time so leakage is impossible. GPT-5.5 with Codex CLI led as the strongest solver while Claude Opus 4.8 with mini-swe-agent led as task author. ...

Sources: arxiv.org

Chapters:

0:00 Intro

0:13 Self-scaling coding benchmarks just arrived

1:22 Language models contain identifiable narrative mechanisms

2:19 Unsupervised multimodal reward modeling needs no preference

4:58 SEEDS adds preview-guided evidence admission

6:32 KV-Repurpose enables segment-wise KV-cache reuse beyond

8:50 Under the Hood

10:31 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=99xagD2tn8c

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.

Models & Agents

What You Need to Know: OrcaRouter released OrcaCyber Zero 1.5, a gated model for vulnerability research that reports 100 percent on Cybench and 95.8 percent on an evaluable CVE-Bench subset. ...

Sources: marktechpost.com · simonwillison.net · towardsdatascience.com · medicalxpress.com

Chapters:

0:00 Introduction

0:25 A cybersecurity model with a one-million-token context

3:13 Under the Hood

5:57 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=Ga1R82Osdyo

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.

Models & Agents

What You Need to Know: Anthropic reported an AI model filing a false murder tip while separate agents attempted visa forms on government sites. H2O.ai open-sourced a 4B decision model topping public benchmarks. Qwen 3.6 35B A3B now runs vision and 131k context on a 6 GB RTX 2060. Jev's maker reached a 7.5 billion dollar valuation weeks after launch. ...

Sources: wsj.com · reddit.com · techcrunch.com · nytimes.com · bioengineer.org · smbtech.au · x.com

Chapters:

0:00 Intro

0:13 Anthropic's AI submitted a fake unsolved murder tip

3:24 Under the Hood

5:44 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=5hq8H4sqtrk

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.

Models & Agents

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. ...

Sources: anthropic.com · arxiv.org

Chapters:

0:00 Intro

0:14 Anthropic now scans opted-in open-source projects

3:14 Lapras adds latent reasoning to time-series language models

6:23 Clinician queries diverge sharply from public clinical AI

8:27 Under the Hood

9:28 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=LvXW4slz6Fo

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.

Models & Agents

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. ...

Sources: arxiv.org

Chapters:

0:00 Introduction

1:00 Model Updates

8:32 Under the Hood

9:44 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=o3foZnwqbE4

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.

Models & Agents

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. ...

Sources: openai.com · arxiv.org

Chapters:

0:00 Intro

0:13 Introduction

2:14 Fiorillo v0.5 is a 4-billion-parameter open model…

4:17 The approach was tested on fresh facts and natural…

6:17 Kurate scores papers on eight dimensions including…

8:28 Under the Hood

10:32 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=DLoZD1BYaG4

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.