> downloaded Claude Code > thought it was just another AI chatbot > opened it up > found a chat bar > typed a prompt > nothing happened to my files > realized the difference > browser AI waits for you to bring your files > Claude Code goes into your folder > all 45 PDFs > all the transcripts > all the messy Excel sheets > reads them all > remembers them session after session > you never re-explain your project again > so before your next literature review, do this: > create a folder with all your papers > open it in Claude Code > write a CLAUDEmd with your role, field, and writing style > ask it to screen every PDF using your inclusion/exclusion criteria > let it rename, extract, and summarize while you think > the AI didn't replace your scholarship > it eliminated the 6 hours of grunt work before it starts

Stop guessing whether your agent is getting better. LangChain's co-founder shows how to compare agent setups and use recorded runs to find problems. 03:44 - hooks around model and tool calls 06:55 - when to build a custom harness 12:29 - agent accuracy, latency, and cost 13:04 - debugging failed agent runs 17:29 - suggested fixes from agent traces This 24-min video is worth watching before you spend another week fixing the wrong part of your agent. Watch this first, then read the guide below on building a harness with Jev.

Boris Cherny: "If Claude has a target to iterate against, it can do much better." He shows how tests and screenshots let Claude inspect its own output. In 18 mins, Cherny covers tools, planning, tests, and project instructions. Spend your time on this before spending a f*cking fortune on another course. Watch his explanation, then use the article below to set up your weekly review.

Diogo Almeida, TypeSafe CEO: "You get what you optimize for." His 36-minute talk traces the gap between helpful chat and reliable automation. This gives you a clearer understanding of training goals than many paid AI courses. Watch the talk, then use Together's guide to train your own classifier for about $17 in training costs.

Stop paying premium prices for every coding task. In under 16 mins, DigitalOcean engineers show how to route coding tasks to different models and check what happens to quality, speed, and cost. 02:48 - small models vs frontier models 04:51 - a separate model for routing 06:57 - models for bug fixes, code, and tests 09:07 - router vs Opus: quality and speed 09:41 - OpenCode demo: app, tests, and README I've seen $500 courses that teach less than this. Watch this first, then read article below on experimenting with cheaper models in Codex and choosing the right model for each task.

Greg Brockman (OpenAI president): "Astra has really hit something that I'm like, okay, I think this is pretty reasonable to call it AGI." He explains what Astra can do and why the right tools and context matter for getting useful work done. This helps you decide what to hand over to AI and what you still need to check yourself. Watch the video, then use the article below to split a project into clear tasks with checks for completion.

Diogo Almeida (Jev creator): "System messages are like disgusting global variables." He explains how to split AI workflows into decisions you can test separately. This one video is more useful than many AI courses people pay for. Watch the breakdown, then use the article's prompts to test Jev's relevance checks on videos from your niche.

LangChain co-founder Harrison Chase: "There's a harness that orchestrates a model and some context." In under 24 minutes, he explains the loop behind tool use and context. That helps you understand what changes when you switch runtimes. I'd put this free talk ahead of many paid courses on agent basics. Watch the talk, then read the article to build products that work with different agent harnesses.

Stop explaining your preferences to your AI agent over and over. LangChain's co-founder shows how to turn user feedback into saved instructions an AI agent can use next time. 06:52 - immediate and background memory updates 31:27 - tools for saving and searching memories 46:12 - correcting an outdated memory 53:00 - sorting emails using past examples 1:11:54 - updating agent instructions from feedback A workshop like this shouldn't be available for free. Watch the workshop first, then read the article below for a closer look at Instinct's memory.

Stop paying for AI automation courses until you've watched this. The founder of TypeSafe explains why AI gets things wrong and what it needs to work reliably. 04:29 - assistance and automation 06:15 - what RLHF actually rewards 08:22 - why wrong answers sound convincing 11:54 - automating small, repetitive tasks 16:19 - calibrated decisions for software I've seen $500 courses that teach less than this. Watch this, then read the article below on how to use Jev correctly.

Cormac Brick (Tech Lead at Google AI Edge): “So you can get really robust and reliable function calling using this fine-tuning workflow.” He explains how fine-tuning helped a small AI model reach over 90% success on specific app tasks. Watch the talk, then use the article below to train on your own reviewed examples.

GPT-4 co-author Diogo Almeida: "We are doing a third thing that is optimized for calibrated decision-making" In 18 minutes, he explains why models trained for human approval can be overconfident. You’ll get more out of this video than you would from a $700 course. Watch the talk, then read the guide below to learn what Jev is good at and how to use it.

Andrej Karpathy: "The hottest new programming language is English." In under 40 minutes, he explains how LLMs have revolutionized what we can do in the digital environment and how to use them effectively. This talk contains much more useful info than the $500 course. Watch the full talk, then read the article below to set up a working harness for your agents.

a plain yes-or-no decision inside agent can cost up to 400x more than it has to Harrison Chase, LangChain's CEO, spent 47 mins on what's actually inside an agent's harness 08:27 - what actually counts as a harness 18:14 - the building blocks of an agent 29:51 - why agents need a sandbox 32:37 - sandboxes against prompt injection 41:51 - the mess of evaluating agents watch this first, then read the article below on the model built to handle that decision for a fraction of the price

Andrej Karpathy: "You can outsource your thinking, but you can't outsource your understanding." In this 29-minute talk, he explains why an agent still needs a person directing it. It reveals more useful tips than most paid courses about agents. Watch the talk first, then read the article below for the method that puts it into practice.

You can feed your AI everything you know and it will still ignore most of it. An Anthropic engineer spent 2h live building an agent, showing exactly why a full folder means nothing without context engineering. 05:15 - what a harness is made of: tools, prompts, skills 25:20 - the agent loop: act, observe, repeat 49:30 - read-only vs read-write file permissions 58:15 - context engineering with ls and cat 1:21:20 - building a multi-step research agent This workshop shouldn't have been given to the public for free. Watch this first, then read the article below on how to feed your AI the info so it actually uses it.

the founder of @videoclaw quit his developer job and moved back in with his parents to build his own products. then came the videos needed to market them. he bought recording gear, paid editors and tried a bunch of AI tools. vibe coding was fast. video creation was still slow. eventually, he built Videoclaw. its desktop video creation agent is now launching in public beta. i’d start by asking it to cut a long product demo into something people might actually finish watching. then work through the changes in chat, much like vibe coding. the AI editor, video generator, avatar maker and voice tools are all in the same app. AI automates complex production work, and simple requests let you make precise edits. you can ask it to fix a caption without hunting through menus yourself. that’s useful for developers and founders trying to explain what they’ve built. marketers have plenty to make with it, too. the rest of us get a way to speak video without needing the time, skills or team to handle production ourselves. acrylic paints, Photoshop, Canva, AI image generation. easier tools keep bringing more people into creating. Videoclaw lowers that barrier for video, and i’m curious what people who’ve never made one will do with it. maybe the next generation of creators ends up calling themselves content engineers. seems plausible when so much of the work happens through instructions. the public beta is free to use, with $10 in AI generation credits included. the first 300 people get $50. it’s a Mac app for now, with more platforms coming soon. download it at videoclaw.com, connect your ChatGPT or Claude, and start prompting.

Stop routing every agent call through your most expensive model. Engineer at Uber explains how to route agent calls to the right model. 00:00 - 70% of PRs now come from agents 01:26 - one gateway routes every model call 03:51 - cutting the token tax by 40% 08:39 - one context graph instead of 20 scattered systems 17:29 - the bottleneck is now deciding what to build I've seen $500 courses that teach less than this. Watch this first, then read the article below on running your own agents for less, without reducing your results.

Your code can route support tickets and send low-confidence cases to a human Jev is @typesafeai's first public System One Model, built for fast, cheap decisions inside software For a support ticket, Jev can assess urgency, check whether it concerns billing and identify whether a human is needed, all in parallel in one request It returns typed answers with probabilities, so you can combine decisions in code and set confidence thresholds for automatic action or review At $0.042 per million input tokens, with free output, I'd use the same approach for form follow-ups and deciding which notifications are worth sending Read @CompleteSkeptic's launch post and join Jev's newly opened developer waitlist

Odyssey-3 genuinely surprised me. one world model that powers a humanoid, controls robot arms, drives a car, pilots a drone, plays video games and trains other AIs. the team behind @odysseyml built self-driving cars at Cruise and Wayve and has been on world models since 2023. so when they call task-by-task robot training brute force, I listen. their answer is a physics agent. it learns how the world works by watching it: physics, motion, cause and effect, how people behave. a new machine plugs into that knowledge. they say a few hours of its own data gets it going, with an action head turning the model's understanding into controls. it's basically what LLMs did for knowledge work. one broad model underneath, and every upgrade to it makes everything built on top better. what stuck with me most: it also generates worlds where AI agents can practice, so their mistakes happen in a simulation first. they think world models will end up understanding physics better than we do. big claim. I want to see it hold up when it's public in the coming weeks.

I gave the same prompt to GLM 5.3 Flash, Kimi K3 and DeepSeek v4 Pro and got 3 working apps back. Built all 3 inside @boltdotnew and shipped the best one to a live link. On a normal plan I'd never burn prompts on a comparison like that. Forge is an experimental mode in Bolt that runs open models with up to 50x more usage. You can build and iterate on web apps without worrying about usage limits. One monthly bar, no daily cap. Draft 10 versions if you need to, then switch to Standard or Max when it's time to ship. On Bolt's own build benchmark, the Forge models score 91% of the top paid model. Close enough for drafting. The deal behind it: opt in, and your builds help train open models. It asks every time you switch in. The research preview runs Sept 14 to Oct 14, 2026. Forge is free on Pro plans until Oct 14, and there's a $9/month early-access plan, Bolt Lite, with a waitlist. It's experimental, so duplicate your project before bringing anything serious in. Watch the 3 builds, then try Forge with the link below.