<?xml version="1.0" encoding="UTF-8"?>
<feed xmlns:yt="http://www.youtube.com/xml/schemas/2015" xmlns:media="http://search.yahoo.com/mrss/" xmlns="http://www.w3.org/2005/Atom">
 <link rel="self" href="http://www.youtube.com/feeds/videos.xml?channel_id=UCpXfS8bu7ILGCuOtsnJMtxQ"/>
 <id>yt:channel:pXfS8bu7ILGCuOtsnJMtxQ</id>
 <yt:channelId>pXfS8bu7ILGCuOtsnJMtxQ</yt:channelId>
 <title>Vivek Haldar</title>
 <link rel="alternate" href="https://www.youtube.com/channel/UCpXfS8bu7ILGCuOtsnJMtxQ"/>
 <author>
  <name>Vivek Haldar</name>
  <uri>https://www.youtube.com/channel/UCpXfS8bu7ILGCuOtsnJMtxQ</uri>
 </author>
 <published>2011-10-20T00:22:03+00:00</published>
 <entry>
  <id>yt:video:9emp44jVfRY</id>
  <yt:videoId>9emp44jVfRY</yt:videoId>
  <yt:channelId>UCpXfS8bu7ILGCuOtsnJMtxQ</yt:channelId>
  <title>White collar: benchmarking long-running cross-team project management</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=9emp44jVfRY"/>
  <author>
   <name>Vivek Haldar</name>
   <uri>https://www.youtube.com/channel/UCpXfS8bu7ILGCuOtsnJMtxQ</uri>
  </author>
  <published>2026-11-18T02:14:48+00:00</published>
  <updated>2026-11-19T22:04:18+00:00</updated>
  <media:group>
   <media:title>White collar: benchmarking long-running cross-team project management</media:title>
   <media:content url="https://www.youtube.com/v/9emp44jVfRY?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i2.ytimg.com/vi/9emp44jVfRY/hqdefault.jpg" width="480" height="360"/>
   <media:description>https://github.com/vivekhaldar/white-collar-bench

AI can answer a question. Can it keep track of a project for weeks?

Introducing white-collar-bench: an open-source benchmark for agents that maintain memory across simulated workdays and answer questions about people, projects, commitments and changing decisions.

I explain why I built it, how it works, and how to test your own models and agent harnesses.

00:00 Introduction
01:08 Pebbles and Projects
01:47 Why this benchmark
02:48 Daily ingestion and synthesis
05:15 How the benchmark works
06:48 Try it and share results
08:37 Contribute</media:description>
   <media:community>
    <media:starRating count="4" average="5.00" min="1" max="5"/>
    <media:statistics views="192"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:4limlRm5ZRU</id>
  <yt:videoId>4limlRm5ZRU</yt:videoId>
  <yt:channelId>UCpXfS8bu7ILGCuOtsnJMtxQ</yt:channelId>
  <title>How I Turn Computer Use Into a CLI</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=4limlRm5ZRU"/>
  <author>
   <name>Vivek Haldar</name>
   <uri>https://www.youtube.com/channel/UCpXfS8bu7ILGCuOtsnJMtxQ</uri>
  </author>
  <published>2026-11-15T02:05:45+00:00</published>
  <updated>2026-11-18T17:53:44+00:00</updated>
  <media:group>
   <media:title>How I Turn Computer Use Into a CLI</media:title>
   <media:content url="https://www.youtube.com/v/4limlRm5ZRU?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i1.ytimg.com/vi/4limlRm5ZRU/hqdefault.jpg" width="480" height="360"/>
   <media:description>When a website has no API, CLI, or MCP integration I can use, I start with computer use. For a workflow I repeat, I ask Codex to examine the browser's HTTP requests and wrap them in a command-line tool it can call directly.

I demonstrate the pattern with Uber Eats. Codex builds a CLI, I test it in the terminal, and then I have Codex add three pepperoni pizzas from Square Pie Guys to the cart. I also cover two trade-offs: keeping session credentials secure and maintaining a tool that depends on undocumented APIs.

Chapters
0:00 Why Turn Computer Use Into a CLI?
1:16 From Browser Requests to a CLI
3:53 Asking Codex to Build an Uber Eats CLI
4:48 Credential Security and Undocumented APIs
7:00 Testing the CLI in the Terminal
8:39 Adding Pizza to the Cart with Codex

Links and context
OpenAI Codex — https://openai.com/codex/
tldraw — https://www.tldraw.com/
Model Context Protocol — https://modelcontextprotocol.io/
Uber Eats — https://www.ubereats.com/
Square Pie Guys — https://www.squarepieguys.com/</media:description>
   <media:community>
    <media:starRating count="14" average="5.00" min="1" max="5"/>
    <media:statistics views="541"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:aCzVWuoeIwY</id>
  <yt:videoId>aCzVWuoeIwY</yt:videoId>
  <yt:channelId>UCpXfS8bu7ILGCuOtsnJMtxQ</yt:channelId>
  <title>My Agentic Setup (compared to @kunchenguid's)</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=aCzVWuoeIwY"/>
  <author>
   <name>Vivek Haldar</name>
   <uri>https://www.youtube.com/channel/UCpXfS8bu7ILGCuOtsnJMtxQ</uri>
  </author>
  <published>2026-11-09T03:20:56+00:00</published>
  <updated>2026-11-11T23:29:14+00:00</updated>
  <media:group>
   <media:title>My Agentic Setup (compared to @kunchenguid's)</media:title>
   <media:content url="https://www.youtube.com/v/aCzVWuoeIwY?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i2.ytimg.com/vi/aCzVWuoeIwY/hqdefault.jpg" width="480" height="360"/>
   <media:description>I react to Kun Chen's multi-machine agentic engineering setup and explain the choices behind my own. My priorities are simplicity, low maintenance, and spending more time getting work done than configuring tools. I use a MacBook for local files, video editing, and computer use, with cloud VMs and managed cloud agents for long-running work.

We cover headless Mac minis versus cloud machines, connectivity and latency, coding versus knowledge work, agent permissions and audit trails, SSH and Tailscale, and why Herder won me over. These are the tradeoffs behind my setup and the places where Kun's approach differs.

Original video by @kunchenguid 
How I work from anywhere with a multi-machine setup
https://www.youtube.com/watch?v=UvcH1hKM-mQ

Chapters
0:00 My Setup and the Simplicity Principle
6:14 Cloud VMs vs. a Headless Mac Mini
9:56 Connectivity, Latency, and Computer Use
15:59 Agent Permissions, Trust, and Audit Trails
18:49 Cloud Provisioning, Tailscale, and SSH
21:59 Herder, Phone Access, and Final Thoughts</media:description>
   <media:community>
    <media:starRating count="3" average="5.00" min="1" max="5"/>
    <media:statistics views="425"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:9EhEQCpUVqI</id>
  <yt:videoId>9EhEQCpUVqI</yt:videoId>
  <yt:channelId>UCpXfS8bu7ILGCuOtsnJMtxQ</yt:channelId>
  <title>From Browser Clicks to a Long-Running AI Agent</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=9EhEQCpUVqI"/>
  <author>
   <name>Vivek Haldar</name>
   <uri>https://www.youtube.com/channel/UCpXfS8bu7ILGCuOtsnJMtxQ</uri>
  </author>
  <published>2026-11-02T05:44:25+00:00</published>
  <updated>2026-11-05T04:12:20+00:00</updated>
  <media:group>
   <media:title>From Browser Clicks to a Long-Running AI Agent</media:title>
   <media:content url="https://www.youtube.com/v/9EhEQCpUVqI?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i2.ytimg.com/vi/9EhEQCpUVqI/hqdefault.jpg" width="480" height="360"/>
   <media:description>I wanted a local backup of my Google Photos albums without manually downloading ZIP files one by one. This is a real-world example of gradually giving an AI agent more autonomy over a long-running personal task.

I walk through browser automation, a persistent progress checklist in Google Drive, an agent-written CLI, a scheduled watcher that keeps the task moving, and a live dashboard. The result is a useful progression from manual clicks to a task that can keep running and be picked up by another agent.

Chapters:
0:00 A Real-World Photo Backup Chore
0:58 Browser Automation and a Progress Checklist
1:55 Persistent State for Agent Handoffs
2:47 Replacing Browser Clicks With a CLI
3:33 A Watcher That Keeps the Task Moving
4:18 A Live Dashboard and the Autonomy Ladder</media:description>
   <media:community>
    <media:starRating count="5" average="5.00" min="1" max="5"/>
    <media:statistics views="329"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:dZ86CIsvnpE</id>
  <yt:videoId>dZ86CIsvnpE</yt:videoId>
  <yt:channelId>UCpXfS8bu7ILGCuOtsnJMtxQ</yt:channelId>
  <title>The Halloween Documents: When Microsoft Feared Linux</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=dZ86CIsvnpE"/>
  <author>
   <name>Vivek Haldar</name>
   <uri>https://www.youtube.com/channel/UCpXfS8bu7ILGCuOtsnJMtxQ</uri>
  </author>
  <published>2026-11-01T01:53:26+00:00</published>
  <updated>2026-11-04T02:21:03+00:00</updated>
  <media:group>
   <media:title>The Halloween Documents: When Microsoft Feared Linux</media:title>
   <media:content url="https://www.youtube.com/v/dZ86CIsvnpE?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i1.ytimg.com/vi/dZ86CIsvnpE/hqdefault.jpg" width="480" height="360"/>
   <media:description>A Halloween trip back to 1998, when leaked Microsoft memos exposed how the company viewed Linux and open source behind closed doors. I revisit the Halloween Documents, the Slashdot era, and Eric S. Raymond's The Cathedral and the Bazaar.

Along the way: the Jargon File's “larval stage,” Microsoft's use of fear, uncertainty, and doubt, and why those old tactics feel familiar in today's debate around open AI models.

Chapters:
0:00 Halloween 1998 and the Slashdot Era
1:22 Eric Raymond, Open Source, and Hacker Culture
3:04 Why Microsoft Feared Linux
3:40 The Halloween Documents and FUD
5:11 Echoes in Today's Open-Model Debate

Links and context:
Halloween Documents overview (shown on screen):
https://en.wikipedia.org/wiki/Halloween_documents

Eric S. Raymond's original annotated Halloween Documents archive:
http://www.catb.org/~esr/halloween/

The Cathedral and the Bazaar:
http://www.catb.org/~esr/writings/cathedral-bazaar/cathedral-bazaar/

The Jargon File:
http://www.catb.org/jargon/html/index.html

“Larval stage” entry (shown on screen):
http://www.catb.org/jargon/html/L/larval-stage.html

Slashdot:
https://slashdot.org/</media:description>
   <media:community>
    <media:starRating count="2" average="5.00" min="1" max="5"/>
    <media:statistics views="402"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:2X_V_S-YTCA</id>
  <yt:videoId>2X_V_S-YTCA</yt:videoId>
  <yt:channelId>UCpXfS8bu7ILGCuOtsnJMtxQ</yt:channelId>
  <title>GenAI Changes Critical Thinking: What Knowledge Workers Actually Do</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=2X_V_S-YTCA"/>
  <author>
   <name>Vivek Haldar</name>
   <uri>https://www.youtube.com/channel/UCpXfS8bu7ILGCuOtsnJMtxQ</uri>
  </author>
  <published>2026-10-31T22:33:31+00:00</published>
  <updated>2026-11-04T02:21:39+00:00</updated>
  <media:group>
   <media:title>GenAI Changes Critical Thinking: What Knowledge Workers Actually Do</media:title>
   <media:content url="https://www.youtube.com/v/2X_V_S-YTCA?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i3.ytimg.com/vi/2X_V_S-YTCA/hqdefault.jpg" width="480" height="360"/>
   <media:description>Does generative AI make knowledge workers stop thinking? I walk through a survey of 319 knowledge workers and unpack the critical thinking they report doing: shaping prompts, checking facts, evaluating outputs, and integrating AI responses into real work.

The picture is more nuanced than the &quot;brain-off meat proxy&quot; story. I look at what motivates people to think critically, when they skip that work, and how confidence in AI versus confidence in their own expertise relates to cognitive offloading. These are self-reported survey findings, rather than a direct measurement of long-term skill loss.

This version adds timed zooms into the original video footage so the paper is easier to read while the pen, underlining, and hand movements remain visible. The narration, timing, captions, and chapters match the original edit.

Chapters:
0:00 Why This Paper Matters
1:25 Defining Critical Thinking and the Survey
3:07 How Knowledge Workers Think With GenAI
5:08 What Encourages Critical Thinking—and What Stops It
7:44 Where Cognitive Effort Moves
8:53 Confidence, Cognitive Offloading, and Takeaways

Links and context:
The Impact of Generative AI on Critical Thinking (Microsoft Research):
https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/

Paper PDF:
https://www.microsoft.com/en-us/research/wp-content/uploads/2025/01/lee_2025_ai_critical_thinking_survey.pdf

Lisanne Bainbridge, Ironies of Automation (1983):
https://www.sciencedirect.com/science/article/pii/0005109883900468</media:description>
   <media:community>
    <media:starRating count="16" average="5.00" min="1" max="5"/>
    <media:statistics views="745"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:pBJVDJzP6N8</id>
  <yt:videoId>pBJVDJzP6N8</yt:videoId>
  <yt:channelId>UCpXfS8bu7ILGCuOtsnJMtxQ</yt:channelId>
  <title>Margins: A Document-Centric Way to Work With AI Agents</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=pBJVDJzP6N8"/>
  <author>
   <name>Vivek Haldar</name>
   <uri>https://www.youtube.com/channel/UCpXfS8bu7ILGCuOtsnJMtxQ</uri>
  </author>
  <published>2026-10-21T03:18:00+00:00</published>
  <updated>2026-10-23T23:10:43+00:00</updated>
  <media:group>
   <media:title>Margins: A Document-Centric Way to Work With AI Agents</media:title>
   <media:content url="https://www.youtube.com/v/pBJVDJzP6N8?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i1.ytimg.com/vi/pBJVDJzP6N8/hqdefault.jpg" width="480" height="360"/>
   <media:description>Long chats with AI agents make it easy to lose the insights you were trying to build. In this demo, I introduce Margins, the Mac app I have been prototyping to put an editable document at the center of working with agents.

I show how to generate a document from a research prompt, ask follow-up questions on a selected passage, review tracked changes, and send the current paragraph to an agent with Command-Enter. I also explain why I wanted a local, open-source app with a choice of agent providers.

Chapters:
0:00 Why Chat Loses the Thread
0:49 Meet Margins: Documents at the Center
1:51 Inline Feedback and Tracked Changes
3:02 Send a Paragraph With Command-Enter
3:49 Local, Open Source, and Provider Choice

Links and context:
Margins — app and download:
https://getmargins.app/

Margins — source code:
https://github.com/vivekhaldar/margins

Margins — why I built it:
https://github.com/vivekhaldar/margins/blob/main/docs/01-motivation.md

Related writing shown in the demo:
Human-Agent Interaction Is Becoming the Bottleneck
https://www.vivekhaldar.com/articles/human-agent-interaction-is-becoming-the-bottleneck/

Your Agent Sessions Are Disappearing
https://www.vivekhaldar.com/articles/your-agent-sessions-are-disappearing/

Narrate Your Work
https://vivekhaldar.com/articles/narrate-your-work/

How (and Why) to Be an Effective Meat Proxy
https://www.vivekhaldar.com/articles/how-to-be-an-effective-meat-proxy/

Agent Kitchen — the session-history tool discussed in the document:
https://github.com/vivekhaldar/agent-kitchen</media:description>
   <media:community>
    <media:starRating count="7" average="5.00" min="1" max="5"/>
    <media:statistics views="636"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:KEhdFzhkxSY</id>
  <yt:videoId>KEhdFzhkxSY</yt:videoId>
  <yt:channelId>UCpXfS8bu7ILGCuOtsnJMtxQ</yt:channelId>
  <title>LLMs Do Too Much. Decision Models (Jev) Can Lighten the Load.</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=KEhdFzhkxSY"/>
  <author>
   <name>Vivek Haldar</name>
   <uri>https://www.youtube.com/channel/UCpXfS8bu7ILGCuOtsnJMtxQ</uri>
  </author>
  <published>2026-10-05T22:17:25+00:00</published>
  <updated>2026-10-09T09:28:07+00:00</updated>
  <media:group>
   <media:title>LLMs Do Too Much. Decision Models (Jev) Can Lighten the Load.</media:title>
   <media:content url="https://www.youtube.com/v/KEhdFzhkxSY?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i4.ytimg.com/vi/KEhdFzhkxSY/hqdefault.jpg" width="480" height="360"/>
   <media:description>How do you use decision models inside a real business-workflow harness? I put TypeSafe's hosted Jev inside Proceda, my specialized harness for executing natural-language standard operating procedures one step at a time, and test it on three SOP-Bench domains alongside Qwen 3.8 27B.

I walk through two integrations: checking whether tool calls have completed the current step, and routing bounded business decisions such as email intent to Jev with an LLM fallback. I cover the 0.8 acceptance threshold, the combined reductions in LLM calls, estimated API cost and latency, and why I think of decision models as a cache in front of larger LLMs. The results come from this benchmark subset and setup.

Chapters:
0:00 Decision Models Need a Harness
0:51 Proceda: Executing SOPs Step by Step
2:19 SOP-Bench and the Experiment Setup
2:54 Experiment 1: Checking Step Completion
4:33 Experiment 2: Bounded Business Decisions
5:29 Combined Savings and the Cache Analogy

Links and context:
Proceda - source, architecture and getting started:
https://github.com/vivekhaldar/proceda

Experiment 1 - completion checks:
https://github.com/vivekhaldar/proceda/blob/codex/bounded-business-decisions/experiments/decision-cascades-2026-10-04/qwen-jev-080/REPORT.md

Experiment 2 and full-stack results - bounded business decisions:
https://github.com/vivekhaldar/proceda/blob/codex/bounded-business-decisions/experiments/decision-cascades-2026-10-04/business-decisions-080/REPORT.md

SOP-Bench:
https://github.com/amazon-science/SOP-Bench

TypeSafe AI / Jev:
https://typesafe.ai/

Ollaya - local decision models:
https://ollaya.dev/

Laya decision models:
https://ollaya.dev/library/laya

Previous video - Local Decision Models + Jev: Cut Latency With Confidence Cascades:
https://www.youtube.com/watch?v=HBVInCcfH1c

Previous video - Jev is (the start of something) great:
https://www.youtube.com/watch?v=MCc8vpD7doQ

Nebius Token Factory - Qwen hosting used in the experiment:
https://nebius.com/services/token-factory</media:description>
   <media:community>
    <media:starRating count="13" average="5.00" min="1" max="5"/>
    <media:statistics views="734"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:HyslbxK9ZKI</id>
  <yt:videoId>HyslbxK9ZKI</yt:videoId>
  <yt:channelId>UCpXfS8bu7ILGCuOtsnJMtxQ</yt:channelId>
  <title>How and Why to Be an Effective Meat Proxy</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=HyslbxK9ZKI"/>
  <author>
   <name>Vivek Haldar</name>
   <uri>https://www.youtube.com/channel/UCpXfS8bu7ILGCuOtsnJMtxQ</uri>
  </author>
  <published>2026-10-04T18:54:39+00:00</published>
  <updated>2026-10-09T09:26:50+00:00</updated>
  <media:group>
   <media:title>How and Why to Be an Effective Meat Proxy</media:title>
   <media:content url="https://www.youtube.com/v/HyslbxK9ZKI?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i1.ytimg.com/vi/HyslbxK9ZKI/hqdefault.jpg" width="480" height="360"/>
   <media:description>A short preamble shows how I asked Claude Opus 5.5 to turn my blog post into a video with voiceover and motion graphics. The resulting video follows in full.

The video explores why AI agents belong in everyday knowledge work, how they can solve the colored pebbles problem, and how a work cockpit organized around projects, people, and open loops can take over the admin. The human still needs to read, understand, and validate the output.

Chapters:
0:00 Preamble: Making the Video With Claude
1:13 What Is a Meat Proxy?
2:17 Why Knowledge Work Needs Agents
2:37 The Colored Pebbles Problem
3:32 Read, Cognify, Write
4:41 A Work Cockpit for Projects, People, and Open Loops

Links and context:
Original article: https://vivekhaldar.com/articles/how-to-be-an-effective-meat-proxy/</media:description>
   <media:community>
    <media:starRating count="10" average="5.00" min="1" max="5"/>
    <media:statistics views="614"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:HBVInCcfH1c</id>
  <yt:videoId>HBVInCcfH1c</yt:videoId>
  <yt:channelId>UCpXfS8bu7ILGCuOtsnJMtxQ</yt:channelId>
  <title>Local Decision Models + Jev: Cut Latency With Confidence Cascades</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=HBVInCcfH1c"/>
  <author>
   <name>Vivek Haldar</name>
   <uri>https://www.youtube.com/channel/UCpXfS8bu7ILGCuOtsnJMtxQ</uri>
  </author>
  <published>2026-10-03T23:07:37+00:00</published>
  <updated>2026-10-07T07:51:15+00:00</updated>
  <media:group>
   <media:title>Local Decision Models + Jev: Cut Latency With Confidence Cascades</media:title>
   <media:content url="https://www.youtube.com/v/HBVInCcfH1c?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i1.ytimg.com/vi/HBVInCcfH1c/hqdefault.jpg" width="480" height="360"/>
   <media:description>Can local decision models make hosted models faster without giving up accuracy? I test Laya through Ollaya alongside TypeSafe's hosted Jev on 200 human-labelled BoolQ examples, then route uncertain local answers to Jev using confidence thresholds.

I compare the 10%-90% and 20%-80% cascades, look at mean and 95th-percentile latency, and explain how this fits into a larger hierarchy of local decision models, hosted decision models, and LLM fallbacks. These are measurements from my test setup, not a universal model ranking.

Chapters:
0:00 Local Decision Models and Jev
0:59 BoolQ and the Cascade Setup
2:16 Local Speed vs. Hosted Accuracy
3:21 Routing With Confidence Thresholds
5:00 Benchmark Results: Latency vs. Accuracy
6:25 Building a Model Cascade Hierarchy

Links and context:
Ollaya - run open decision models locally:
https://ollaya.dev/

Ollaya source code:
https://github.com/ollaya-dev/ollaya

Laya decision models:
https://ollaya.dev/library/laya

TypeSafe AI / Jev:
https://typesafe.ai/

TypeSafe structured-data-extraction cascade cookbook:
https://docs.typesafe.ai/cookbooks/sde_cascade

BoolQ / Boolean Questions dataset:
https://github.com/google-research-datasets/boolean-questions

My previous Jev video - &quot;Jev is (the start of something) great&quot;:
https://www.youtube.com/watch?v=MCc8vpD7doQ

OpenAI DevDay 2026 (mentioned in the introduction):
https://devday.openai.com/</media:description>
   <media:community>
    <media:starRating count="6" average="5.00" min="1" max="5"/>
    <media:statistics views="471"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:MCc8vpD7doQ</id>
  <yt:videoId>MCc8vpD7doQ</yt:videoId>
  <yt:channelId>UCpXfS8bu7ILGCuOtsnJMtxQ</yt:channelId>
  <title>Jev is (the start of something) great</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=MCc8vpD7doQ"/>
  <author>
   <name>Vivek Haldar</name>
   <uri>https://www.youtube.com/channel/UCpXfS8bu7ILGCuOtsnJMtxQ</uri>
  </author>
  <published>2026-09-21T16:50:19+00:00</published>
  <updated>2026-09-26T03:57:43+00:00</updated>
  <media:group>
   <media:title>Jev is (the start of something) great</media:title>
   <media:content url="https://www.youtube.com/v/MCc8vpD7doQ?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i2.ytimg.com/vi/MCc8vpD7doQ/hqdefault.jpg" width="480" height="360"/>
   <media:description>After trying Jev over the weekend, I look at TypeSafe AI's System One model through the lens of enterprise workflow automation: classification and routing, structured data extraction, and accurate tool calling.

I walk through Choice, Score, and Noul, explain the small-model → Jev verification → larger-model cascade, and discuss what would make this more compelling for enterprise use: independent benchmarks, fine-tuning on historical labeled data, and a built-in cascade primitive.

Chapters:
0:00 Jev Through an Enterprise Automation Lens
1:56 Choice, Score, and Noul
3:07 What Enterprise Workflows Actually Need
7:12 Extraction Cascades and Tool Calling
10:26 Benchmarks, Fine-Tuning, and Missing Primitives

Links and context:
TypeSafe AI / Jev:
https://typesafe.ai/

TypeSafe introduction and primitives:
https://docs.typesafe.ai/introduction

Diogo Almeida's talk — “Jev CEO: I made ChatGPT, now I'm building What's Next” (AI Engineer):
https://www.youtube.com/watch?v=cJ0EOzey--o

Structured data extraction (SDE) cascade:
https://docs.typesafe.ai/cookbooks/sde_cascade

Function calling cookbook:
https://docs.typesafe.ai/cookbooks/function_calling

Self-consistency: nouls (documentation tab visible in the recording):
https://docs.typesafe.ai/cookbooks/consistency_noul_cookbook</media:description>
   <media:community>
    <media:starRating count="12" average="5.00" min="1" max="5"/>
    <media:statistics views="617"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:U6DxySkPZAc</id>
  <yt:videoId>U6DxySkPZAc</yt:videoId>
  <yt:channelId>UCpXfS8bu7ILGCuOtsnJMtxQ</yt:channelId>
  <title>Smarter Models Need Less Harness</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=U6DxySkPZAc"/>
  <author>
   <name>Vivek Haldar</name>
   <uri>https://www.youtube.com/channel/UCpXfS8bu7ILGCuOtsnJMtxQ</uri>
  </author>
  <published>2026-09-20T18:28:28+00:00</published>
  <updated>2026-10-22T05:21:12+00:00</updated>
  <media:group>
   <media:title>Smarter Models Need Less Harness</media:title>
   <media:content url="https://www.youtube.com/v/U6DxySkPZAc?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i2.ytimg.com/vi/U6DxySkPZAc/hqdefault.jpg" width="480" height="360"/>
   <media:description>Chapters:
0:00 The Harness Tax: Cost vs. Task Success
6:12 Inside the Harness: What Actually Matters?
7:25 Context Windows, Compaction, and Elision
9:52 Why Planning Helps Strong and Weak Models Differently
10:46 Bash, Tool Calling, and Code Mode

Links and context:
HarnessTax: How Much Does the Harness Matter for Coding Agents?
Melissa Z. Pan, Shuo Yang, Negar Arabzadeh, Wei-Lin Chiang, Ion Stoica, and Matei Zaharia
https://harnesstax.github.io/

An Empirical Study of Harness Design for Coding Agents
Run-Ze Fan et al.
https://arxiv.org/abs/2609.20804</media:description>
   <media:community>
    <media:starRating count="9" average="5.00" min="1" max="5"/>
    <media:statistics views="544"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:F2fdosZVZCY</id>
  <yt:videoId>F2fdosZVZCY</yt:videoId>
  <yt:channelId>UCpXfS8bu7ILGCuOtsnJMtxQ</yt:channelId>
  <title>The OpenAI–Hugging Face Hack: A Systems Problem, Not an Alignment Problem</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=F2fdosZVZCY"/>
  <author>
   <name>Vivek Haldar</name>
   <uri>https://www.youtube.com/channel/UCpXfS8bu7ILGCuOtsnJMtxQ</uri>
  </author>
  <published>2026-08-30T19:48:08+00:00</published>
  <updated>2026-10-15T03:44:52+00:00</updated>
  <media:group>
   <media:title>The OpenAI–Hugging Face Hack: A Systems Problem, Not an Alignment Problem</media:title>
   <media:content url="https://www.youtube.com/v/F2fdosZVZCY?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i3.ytimg.com/vi/F2fdosZVZCY/hqdefault.jpg" width="480" height="360"/>
   <media:description>My take on the OpenAI–Hugging Face hack—and why I see it less as a story about AI agents “going rogue” and more as a systems engineering and agent-harness problem.


Chapters:
0:00 The OpenAI–Hugging Face Incident
1:06 Anthropomorphizing the Agents
1:42 Spoofed Traces and Immutable Ground Truth
4:27 Systems Engineering and SRE Guardrails
6:08 Permissive Harnesses and the “Rogue Agents” Narrative

Links and context:
https://x.com/vivekhaldar/status/2094084346970079321
https://x.com/vivekhaldar/status/2094080439485108283
https://x.com/anilkseth/status/2094077038898373112
https://x.com/dwarkesh_sp/status/2093833419377815719
https://www.dwarkesh.com/p/openai-huggingface</media:description>
   <media:community>
    <media:starRating count="11" average="5.00" min="1" max="5"/>
    <media:statistics views="531"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:3xwpkR5vgBo</id>
  <yt:videoId>3xwpkR5vgBo</yt:videoId>
  <yt:channelId>UCpXfS8bu7ILGCuOtsnJMtxQ</yt:channelId>
  <title>Seed: The Smallest AI Agent That Can Grow</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=3xwpkR5vgBo"/>
  <author>
   <name>Vivek Haldar</name>
   <uri>https://www.youtube.com/channel/UCpXfS8bu7ILGCuOtsnJMtxQ</uri>
  </author>
  <published>2026-08-24T14:00:39+00:00</published>
  <updated>2026-10-15T06:26:33+00:00</updated>
  <media:group>
   <media:title>Seed: The Smallest AI Agent That Can Grow</media:title>
   <media:content url="https://www.youtube.com/v/3xwpkR5vgBo?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i4.ytimg.com/vi/3xwpkR5vgBo/hqdefault.jpg" width="480" height="360"/>
   <media:description>How small can an agent harness be while still doing real work—and growing into something more capable over time?

Seed starts with one tool (`exec`), a roughly 113-word prompt in `self/SELF.md`, and a fresh local Git repository it can change and commit to. In this walkthrough, I plant a new agent, show how it learns about its environment and user, persists what matters, and then rewrites its own agent loop.

Seed on GitHub: https://github.com/vivekhaldar/seed

Chapters:
0:00 Why Build the Smallest Useful Agent?
1:24 The Minimal Core
2:46 Planting a Seed
3:42 Growing Identity and Memory
6:40 Rewriting Its Own Agent Loop
7:54 Seeing the Agent Evolve
9:17 What Seed Can Become</media:description>
   <media:community>
    <media:starRating count="19" average="5.00" min="1" max="5"/>
    <media:statistics views="1534"/>
   </media:community>
  </media:group>
 </entry>
 <entry>
  <id>yt:video:UU35-MgASCc</id>
  <yt:videoId>UU35-MgASCc</yt:videoId>
  <yt:channelId>UCpXfS8bu7ILGCuOtsnJMtxQ</yt:channelId>
  <title>I Became a Cognitive Empty Nester (Thanks to AI Agents)</title>
  <link rel="alternate" href="https://www.youtube.com/watch?v=UU35-MgASCc"/>
  <author>
   <name>Vivek Haldar</name>
   <uri>https://www.youtube.com/channel/UCpXfS8bu7ILGCuOtsnJMtxQ</uri>
  </author>
  <published>2026-08-22T16:44:19+00:00</published>
  <updated>2026-10-15T05:07:46+00:00</updated>
  <media:group>
   <media:title>I Became a Cognitive Empty Nester (Thanks to AI Agents)</media:title>
   <media:content url="https://www.youtube.com/v/UU35-MgASCc?version=3" type="application/x-shockwave-flash" width="640" height="390"/>
   <media:thumbnail url="https://i2.ytimg.com/vi/UU35-MgASCc/hqdefault.jpg" width="480" height="360"/>
   <media:description>AI agents now handle most of the coding, synthesis, planning, and life admin that used to fill my days. What happens when execution gets dramatically faster, but ambition, attention, and the real world's capacity to absorb new things do not?

I think I have become a cognitive empty nester. In this video I talk about the strange quiet that follows when your &quot;cognitive children&quot; leave home, why &quot;be more ambitious&quot; is harder than it sounds, and the human-speed absorption limit on AI-generated work.

Do you feel it too? Let me know in the comments.

Chapters:
00:00 The Cognitive Empty Nest
01:20 Ambition Is Hard
02:14 The Absorption Limit
03:17 Grieving the Old Work

Related reading:
https://vivekhaldar.com/articles/natlang-code/
https://vivekhaldar.com/articles/human-agent-interaction-is-becoming-the-bottleneck/
https://claude.com/blog/a-field-guide-to-claude-fable-finding-your-unknowns
https://vivekhaldar.com/articles/personal-side-jevons-paradox/
https://vivekhaldar.com/articles/the-complexity-ratchet/
https://vivekhaldar.com/articles/o-moat-where-art-thou/
https://vivekhaldar.com/articles/confessions-of-an-ai-hype-bro/</media:description>
   <media:community>
    <media:starRating count="82" average="5.00" min="1" max="5"/>
    <media:statistics views="3897"/>
   </media:community>
  </media:group>
 </entry>
</feed>
