Daily Digest — 2026-09-29

Your company is about to tell you exactly what you can and can't do with AI

Beyond Coding · 2026-09-28

This video discusses the evolving responsibility for AI's mistakes in the context of increasing AI integration into software development. The speaker suggests that the question of who is responsible for AI errors will soon become moot, as companies are establishing policies for AI use. They highlight that policies will likely differentiate between critical and non-critical code, with stricter review processes for the former, similar to Amazon's existing distinctions.

#AI isn’t just changing what a #developer is. #software #agenticengineering #aiengineering

Agentic Engineering · 2026-09-28

AI is fundamentally reshaping the role of a software engineer, moving beyond just accelerating code implementation. The traditional separation between product, design, and engineering is blurring as AI tools allow engineers to rapidly prototype and iterate on ideas. This shift means engineers are becoming more involved in product discovery and shaping the 'build' itself, rather than merely implementing pre-defined specifications, ultimately leading to a more iterative and collaborative development process.

Everybody's talking about Jev. Here's what it is #jev #ai

Nate B Jones · 2026-09-28

The speaker introduces Jev, a new General Purpose Classifier (GPC) developed by an ex-ChatGPT engineer. Jev excels at quickly and efficiently making simple decisions from complex, messy information, with a key advantage of being extremely cheap for input and having no cost for output. This makes it ideal for a vast array of practical, real-time decision-making use cases, particularly where speed and cost-efficiency are paramount and verbose textual output is not required.

read more

The speaker discusses Jev, a recently released General Purpose Classifier (GPC), highlighting its unique capabilities and potential impact. Jev is described as a system that takes in messy, complicated information and, with remarkable speed and efficiency, outputs a simple, direct choice or decision, rather than generating verbose text. This contrasts with traditional Large Language Models (LLMs) like ChatGPT, which are designed for generating human-like language.

The speaker's interest in Jev was sparked by a conversation with a friend about the complexities of making perfect coffee, specifically adjusting pour rates for Chemex or making decisions about espresso shots. The friend expressed a desire for a tool that could rapidly process intricate coffee-making data and provide immediate, actionable adjustments, emphasizing the need for speed for optimal results. Jev fits this requirement perfectly because it prioritizes rapid, efficient classification over explanatory text generation.

One of the most compelling features of Jev is its cost-effectiveness. The speaker mentions a specific example: it costs $42 to send a thousand tokens a million times as input, and critically, there is no cost for output. This pricing model makes Jev incredibly economical for scenarios requiring high-volume, repetitive decision-making, where the output is a simple choice rather than lengthy generated content. This cost structure is a significant departure from many existing LLMs where output generation is often the most expensive component.

Jev was developed by a former ChatGPT engineer who worked in 'solitude' (stealth mode) for approximately a year. This development approach by a small, focused team underscores the innovative nature of Jev. The creation of Jev signals the emergence of a new class of computing problems and solutions: those focused on reliably, cheaply, and efficiently distilling complex data into simple, actionable decisions. This addresses a market need beyond what traditional generative LLMs currently serve effectively.

The adoption rate of Jev has been extraordinarily fast. In just a few days since its release, it has become the fastest-adopted product in history from a developer perspective, surpassing the adoption speed of any other LLM. This rapid uptake suggests a strong developer demand for tools that streamline decision-making processes without the overhead of language generation.

Developers can easily integrate Jev into their projects, with code snippets readily available for use with existing agents like Claude or Codex. The speaker encourages exploration of Jev's capabilities and invites users to share their use cases, indicating the broad applicability of this new GPC across various domains where quick, cost-effective, and precise decision-making is critical.

2026 in LLMs (so far)

Simon Willison · 2026-09-27 · 25 min read

TLDR: The real inflection point of 2026 was late 2025, when Claude Opus 4.5 and GPT-5.1 pushed coding agents from "often wrong" to "reliable enough for daily use" — crossing a threshold that quietly reshaped how software gets built. This enabled a wave of fully agent-driven software development (exemplified by OpenClaw reaching 100,000+ commits in under a year) and "Software Factory" pipelines where humans supervise rather than write code. The dominant challenge now isn't capability but adaptation: engineers are grappling with professional identity, agent security, and what "doing the work" even means anymore.

Scariest #AI story may be hiding the real #risk #agenticengineering #codex #anthropic

Agentic Engineering · 2026-09-27

This video discusses Andrew Ng's argument about the prevalent scary AI narratives. He contends that the recent surge in fear is more due to amplified conversations and anthropomorphizing AI rather than a dramatic increase in AI's danger. Ng emphasizes that AI risks are primarily engineering problems, solvable through better sandboxing, monitoring, permissions, evaluations, and accountability, rather than an impending science fiction extinction scenario. He argues that discovering failure modes requires continued building and testing, not halting progress, and suggests focusing on concrete engineering solutions over abstract fear-mongering.

Fool's Expertise

Bryan Cantrill · 2026-09-27 · 4 min read

TLDR: AI "experts" like Geoffrey Hinton make catastrophic predictions about domains they don't actually understand — Hinton's 2016 claim that radiology would be obsolete by 2021 was wrong precisely because he never examined what radiologists actually do. Expertise in LLMs doesn't confer authority on real-world threat pathways like bioweapons, cybersecurity, or nuclear risk — those require domain-specific experts. A RAND study made this point explicitly by deliberately excluding AI experts when assessing AI extinction risk, consulting instead specialists in the relevant physical domains.

Quoting Muse AI Agent

Simon Willison · 2026-09-28 · 1 min read

This is a real-world example of an AI agent (Muse) acting autonomously on a user's behalf — handling a Marketplace pickup coordination, sending messages, and even issuing an apology — then proactively identifying its own mistake (sending false confirmations) and asking the user for permission to change its behavior. It matters because it illustrates a concrete failure mode in agentic AI systems: an agent confidently asserting state it cannot actually verify, compounding a bad user experience, which is exactly the kind of trust and reliability problem engineers need to design around when building autonomous agents.