Recursive language models and small-model performance boost

Andy covers MIT's recursive-model technique that lets smaller models self-call to double long-context reasoning performance.

Play episode from 25:43

chevron_right

Transcript

chevron_right

Transcript

Episode notes

The October 16th episode opened with Brian, Beth, Andy, and Karl discussing the latest AI headlines — from Apple’s new M5 chip and Vision Pro update to Anthropic’s Haiku 4.5 release. The team also broke down a new tool called Hux and explored how managers may be unintentionally holding back their employees’ AI potential.

Key Points Discussed

She Leads AI Conference: Beth shared highlights from the in-person event and announced a virtual version coming November 10–11 for international audiences.

Anthropic’s Haiku 4.5 Launch: The new model beats Sonnet 4 on benchmarks and introduces task-splitting between models for cheaper, faster performance.

Apple’s M5 Chip: The new M5 integrates CPU, GPU, and neural processors into MacBooks, iPads, and a final version of the Vision Pro. Apple may now pivot toward AI-enabled AR glasses instead of full VR headsets.

OpenAI x Salesforce Integration: Karl covered OpenAI’s new deep link into Salesforce, giving users direct CRM access from ChatGPT and Slack. The team debated whether this “AI App Store” model will succeed where plugins and Custom GPTs failed.

Google Gemini 3.1 & Flow Upgrade: Brian demoed the new Flow video engine, which now supports longer, more consistent shots and improved editing precision. The panel noted that consistency across scenes remains the last hurdle for true AI filmmaking.

OpenAI Sora Updates: Pro users can now create 25-second videos with storyboard tools — pushing generative video closer to full short-form storytelling.

Creative AI Discussion: The hosts compared AI perfection to human imperfection, noting that emotion, flaws, and authenticity still define what connects audiences.

MIT Recursive Language Models: Andy shared news of a new technique allowing smaller models to outperform large ones by reasoning recursively — doubling performance on long-context tasks.

Tool of the Day – Hux:

Built by the original NotebookLM team, Hux is an audio-first AI assistant that summarizes calendar events, inboxes, and news into short daily briefings.

Users can interrupt mid-summary to ask follow-ups or request more technical detail.

The team praised Hux as one of the few AI tools that feels ready for everyday use.

Main Topic – Managers Are Killing AI Growth:

Based on a video by Nate Jones, the team discussed how managers who delay AI adoption may be stunting their teams’ career growth.

Karl argued that companies still treat AI budgets like software budgets, missing the need for ongoing investment in training and experimentation.

Andy emphasized that employees in companies that block AI access will quickly fall behind competitors who embrace it.

Brian noted clients now see value in long-term AI partnerships rather than one-off projects, building training and development directly into 2026 budgets.

Beth reminded listeners that this is not traditional “software training” — each model iteration requires learning from scratch.

The panel agreed companies should allocate $3K–$4K per employee annually for AI literacy and tool access instead of treating it as a one-time expense.

Timestamps & Topics

00:00:00 💡 Intro and show overview

00:01:34 🎤 She Leads AI conference recap

00:03:42 🤖 Anthropic Haiku 4.5 release and pricing

00:04:49 🍏 Apple’s M5 chip and Vision Pro update

00:09:03 ⚙️ OpenAI and Salesforce integration

00:16:16 🎥 Google Gemini 3.1 Flow video engine

00:21:11 🧠 Consistency in AI-generated video

00:23:01 🎶 Imperfection and human creativity

00:25:55 🧩 MIT recursive models and small model power

00:28:21 🎧 Hux app demo and review

00:36:35 🧠 Custom AI workflows and use cases