Andrew Ng:
"100% of my tasks now run through AI agents, the hype actually passed my expectations, loops are the next step"
"In 3-6 months, everyone will be using self-improving loops, prompting alone is over"
In a 30-minute talk, Andrew Ng breaks down how to build self-improving agentic systems from scratch
Agent loops
Feedback
Memory
Self-improvement
Worth more than most $500 agentic courses
I genuinely don't understand why everyone isn't using this yet.
Andrej Karpathy, OpenAI co-founder, posted a simple idea that went massively viral:
Stop using AI to write code.
Use it to build a second brain.
You point Claude Code at a folder. Drop in any source: an article, a transcript, a PDF.
Claude reads it, links it, files it into a living wiki of everything you know.
It compounds like interest. The more you feed it, the smarter it gets.
Here's the whole thing:
1) Install Obsidian
2) Create a vault
3) Open it in Claude Code
4) Paste Karpathy's wiki idea and tell Claude to build it
5) Claude makes three folders:
- raw (for sources)
- wiki (for its pages)
- CLAUDE. md (that runs it)
6) Drop any source into raw and say: "ingest this"
7) Ask questions across everything, forever
Five minutes to set up and you never start from a blank chat again.
Full step by step guide below.
THE SYSTEM BUILT TO PREVENT ANOTHER 2008 CRASH JUST FAILED.
After the 2008 financial crisis, regulators rebuilt a type of bond product from the ground up specifically to stop this exact failure from happening again.
They added stricter rules and called the new version CLO 2.0. For over a decade, it worked exactly as designed.
This week, for the first time ever, a CLO built under those post crisis rules actually defaulted.
The fund involved is run by Bain Capital.
Fitch downgraded its riskiest tranche to default after the fund returned €7.4 million to investors on a tranche that was supposed to pay back €11.2 million.
A CLO works by taking a large pool of corporate loans and slicing them into different risk tiers. Investors in the riskiest tier get paid last, but earn the highest return when everything performs well.
This tier did not perform well, and the entire post 2008 safety system did nothing to stop it.
Here is what is actually causing this.
A large share of the loans inside CLOs are tied to software companies.
AI is now directly threatening the business model behind a huge slice of that industry, because AI tools can increasingly do the work that used to require buying and running entire software platforms.
When Anthropic released a major update to its Claude AI model earlier this year, it triggered a real selloff in software company loans, the exact loans many CLOs are built from.
This is not isolated to one fund.
Fitch downgraded three more CLO tranches to triple-C just last month: Barings Euro CLO 2029-2, Man GLG Euro CLO V, and Toro European CLO 6.
JPMorgan estimates that between $40 billion and $150 billion of loans sitting inside CLOs are in sectors directly exposed to AI disruption.
UBS has gone further, warning that in an aggressive AI disruption scenario, default rates across the broader private credit market could climb as high as 13%, compared to roughly 8% for leveraged loans and 4% for high-yield bonds.
There is also a structural problem making this hard to fix from here. Older CLOs like this one have exited what is called their reinvestment period, meaning the manager can no longer swap weak loans for stronger ones.
Refinancing the whole structure is not realistic either, since borrowing costs today are far higher than when these loans were originally written. That leaves only one option: sell off the loan pool and return whatever cash that raises. Nothing more.
JPMorgan CEO Jamie Dimon has already warned that stress in one corner of private credit can signal hidden problems elsewhere, describing it as looking for "cockroaches."
This is the first one confirmed.
I genuinely don't understand why everyone isn't using this yet
Andrej Karpathy, a co-founder of OpenAI, posted a simple idea that hit 16 million views: stop using AI to write code, use it to build a second brain.
You point Claude Code at a folder, drop in any source, an article, a transcript, a PDF, and Claude reads it, links it, and files it into a living wiki of everything you know. It compounds like interest, the more you feed it, the smarter it gets.
Here's the whole thing:
> Install Obsidian, create a vault, open it in Claude Code
> Paste Karpathy's wiki idea file and tell Claude to build it
> Claude makes three folders: raw for sources, wiki for its pages, a CLAUDE.md that runs it
> Drop any source into raw and say "ingest this"
> Ask questions across everything, forever
Five minutes to set up, and you never start from a blank chat again.
Full step-by-step guide with Claude and Obsidian, link below.
Bookmark this
Andrew Ng:
"100% of my tasks are now done by AI agents - hype has exceeded my expectations. Loops is next step.
in 3-6 months, everyone will be using self-improving loops. No more prompting."
In a 30-minute talk, Andrew Ng explains how to build self-improving agentic systems from scratch.
Worth more than a $500 agentic course.
Many people think any given ML project is 99% training.
In reality, it’s 50% evaluation, 40% data cleaning, 8% integration, and 2% training.
The first two set the noise floor for learning. No ML magic matters; the model cannot lower the noise floor, as that’s the optimal bound of Shannon encoding of your data.
Thus, not a single day goes by without me thinking about ontology. Even the old labels have to be constantly reviewed.
I'm a cardiologist. Something just happened today that I genuinely did not see coming — and it could change the future of preventive medicine more than anything I've written about on this platform.
Midjourney — the AI company that became famous for generating images from text prompts — just announced a medical hardware division and unveiled a working prototype of a full-body scanner unlike anything that's ever existed.
It's called the Midjourney Scanner. And it works like this.
You step into a shallow pool of water. You stand on a platform that slowly descends — about two inches per second — through a ring containing roughly half a million tiny ultrasonic transducers, each the size of a grain of sand. Every one of them acts as both a speaker and a microphone, sending ultrasonic waves through your body from every angle and recording what comes back.
60 seconds later, you step out. The scan is done.
No radiation. No magnets. No claustrophobia. No IV contrast. Just sound, water, and an almost incomprehensible amount of computing power — roughly 2 petaflops processing 17 gigabytes per second of raw acoustic data — reconstructing a 3D map of your entire internal anatomy down to half a millimeter resolution.
Organs. Tissues. Blood vessels. Bones. Muscle. Fat distribution. All segmented by AI in real time.
As a cardiologist who has spent months writing about how the standard screening playbook misses the majority of future heart attacks — this is the technology I've been waiting for without knowing it existed.
Here's why this matters for the future of your heart.
Right now, getting a detailed look inside your cardiovascular system requires either a CT scan (radiation), an MRI (magnets, claustrophobia, 45-60 minutes, $1,000+), or a coronary CT angiogram (radiation, IV contrast, limited availability). These are powerful tools. I order them regularly and they save lives.
But they're reactive. You get them when something is already suspected. They're expensive. They're uncomfortable. And for most people, they happen once — maybe twice — in a lifetime.
Imagine instead: a 60-second scan with no radiation that you could repeat monthly or quarterly. Tracking cardiac structure over time. Watching body composition shift. Detecting changes in organ size, fluid distribution, or vascular architecture before symptoms ever develop. Building a longitudinal dataset of YOUR body that AI can analyze for patterns no single snapshot would reveal.
That's what Midjourney is building toward.
The company plans 50,000 scanners worldwide over six years, with capacity for a billion scans per month. The first location — the "Midjourney Spa" in San Francisco — opens at the end of 2027 with 10 scanners alongside saunas, cold plunges, and a gym. The scan costs a few dollars. The experience is designed to feel like wellness, not medicine.
The technology is built on Butterfly Network's ultrasound-on-chip platform — 40 modules per scanner — combined with Midjourney's own AI segmentation and reconstruction stack. David Holz, the founder, claims the system aims for image quality comparable to MRI in many aspects but at nearly 100x the speed with zero radiation.
Now the caveats — because I'm a physician and the caveats matter enormously.
This is a Gen 1 prototype. About a dozen people have been scanned so far. Current scan time is actually closer to 20 minutes, not 60 seconds — the system is bottlenecked by bandwidth and reconstruction algorithms. The 60-second target is aspirational for future hardware generations.
It is not FDA-cleared for diagnostic use. Midjourney is starting with body composition maps — a category below diagnostic imaging in the regulatory hierarchy. The path from "beautiful 3D body scans" to "clinically validated diagnostic tool that your cardiologist can act on" runs through years of clinical trials, comparative studies against MRI and CT gold standards, and FDA review.
No independent clinical validation has been published. The imaging claims come from Midjourney's own demonstrations. Comparative data against established modalities does not yet exist.
And the privacy implications of full-body internal scans at planetary scale — a billion scans per month — is a conversation that hasn't even started yet.
So I want to be precise. This is not ready for clinical medicine today. It may not be ready for years. Many ambitious medical hardware projects have failed in the gap between prototype and product.
But.
The fact that a working prototype exists — producing real segmented 3D anatomy from sound waves and compute alone — means the physics works. The engineering works. The question is no longer "is this possible" but "how fast can it be validated and scaled."
And if it is validated — if the resolution holds up against MRI, if the AI segmentation proves reliable, if the regulatory path clears — then what we're looking at is the most significant new imaging modality in 50 years.
For my entire career, preventive cardiology has been limited by the fact that seeing inside the body is expensive, slow, uncomfortable, and infrequent. We catch disease late because we image rarely. We image rarely because imaging is hard.
A 60-second, no-radiation, spa-based full-body scan that costs a few dollars would demolish every one of those barriers.
I've written about AI detecting inflamed arteries. About gene editing curing cholesterol. About GLP-1 drugs rewriting metabolic medicine. About cellular reprogramming reversing aging.
This is the missing piece: the ability to see inside every human body, routinely, safely, and affordably — so all of those interventions can be deployed before the disease arrives instead of after.
The company that taught AI to generate images from imagination just built a machine that generates images from the human body.
The future of medicine showed up today from the last place anyone expected.
NVIDIA CEO, Jensen Huang:
"Nobody writes prompts anymore. The new job is to write and handle loops."
He calls it the shift that defines the rest of 2026.
Interview was out just yesterday.
Watch the 23 minute talk, then save the full framework below👇
stop telling Claude Code/Codex "do this".
stop telling Claude Code/Codex "write code".
stop telling Claude Code/Codex "fix this bug".
you're using a 𝘀𝗲𝗻𝗶𝗼𝗿 𝗔𝗜 like it's a 𝗷𝘂𝗻𝗶𝗼𝗿 𝗶𝗻𝘁𝗲𝗿𝗻.
here are 8 prompts and goals you can copy-paste directly.
Anthropic posted a FULL GUIDE on how to prompt Fable 5 (Mythos).
Claude Fable 5 is not meant to be prompted like any other model.
It's meant to run autonomously.
Here's exactly how to enable Fable to do work for you with minimal manual intervention:
1. Effort selection
Anthropic recommends using High for most tasks and Xhigh only for complex workflows.
Low/medium: quick questions, basic research
High: default for most work
Xhigh: complex builds, multi-step analysis
Ultracode: full autonomous orchestration
2. /loop prompting
Use /loop prompts to kick Fable off to complete full tasks.
/loop <time interval> + <goal>
3. Tell it WHY, not just what (context)
Fable can't perform on instructions alone. It needs context to make decisions on its own.
Anthropic's exact prompting structure:
"I'm working on [larger task] for [who it's for]. They need [what the output enables]. With that in mind: [your actual request]."
4. Keep prompts short (instructions)
Counterintuitive but critical.
Over-engineering your prompts on Fable 5 degrades output. You're constraining a model that would have figured it out on its own.
4. Tell it when to stop and check in during runs
"Pause for me only when the work genuinely requires my input: a destructive action, a real scope change, or something only I can provide. Otherwise, keep going and report back when done."
5. Build it a memory system
Fable performs best when it can record lessons from its previous loops.
Give it a markdown file and this instruction:
"Store one lesson per file with a one-line summary at the top. Record corrections and confirmed approaches. Don't save what the repo or chat history already records."
The optimal general prompt structure:
"Goal: I'm working on [larger task] for [who it's for]. They need [what the output enables].
Request: [your specific ask in one sentence]
Output format: [exactly how you want it]
Constraints: [what must not happen]."
One last thing - your old prompts may actually work against you.
Skills and project instructions built for Opus 4.8 may produce worse results on Fable.
Bookmark this to actually maximize your Fable workflows.
BREAKING: Claude can now run Stock Market research like a top consulting firm (for free).
Here are 10 Claude prompts that replace $100K/year stock analysts (Save for later)
The new Claude 5 is the smartest AI you ever used.
And you have until June 22 to not pay for it:
1. Claude 5 'Fable' is a new kind of Claude. Understand it's much more than an upgrade.
2. It's the smartest thing you've ever talked to.
3. But Anthropic (Claude's parent company) gives us "free" access to it with the paid plan until the 22nd of June. After this, we'll probably pay, a lot.
4. Now what's so different? Everything, and nothing. You can do the exact same thing as before, but so much better, it feels drastically different.
5. It's slower, though. And it burns tokens (= credits) faster. So don't prompt it "Hi, how are you?"
6. Actually very important: give it your longest, hardest task ever. Aim for the end goal on the first prompt. Ask for the moon and watch it work.
7. On some of my tasks, Claude Fable-5 worked for 20 minutes straight. And it was good. Like coworker-I-should-probably-pay good. Try it.
8. A quick trick: use the "Research" mode (using the + at the bottom left of the chatbox) to ask for a very hard research task. I was seriously impressed.
9. Knowing what to ask is still the most important thing about AI. So yes, prompts matter still, sorry.
10. A trick I've been repeating forever: use the tool "AskUserQuestion" to make Claude ask you questions before answering, so it prompts it better.
11. If you give it good context (who are you, what you love or hate) before giving it a task, it's also much better at answering.
Quick reminder for people who don't know where and how to start, because I wrote about all of this:
✦ https://t.co/jw2qdIcjnh → my "Claude for Dummies".
✦ https://t.co/uWTpOI3Woc → my favorite Claude to work
✦ https://t.co/Vn60ElPZ2i → all of my Claude guides.
📩 Send this post to your colleague who still thinks Claude is the name of an old French guy.
This is a super exciting release - Claude Fable 5 is the same underlying model as Mythos but with added safeguards. The benchmarks are great and it's SOTA on everything by a margin but I'll add that *qualitatively* also, this is a major-version-bump-deserving step change forward (imo of the same order as Claude 4.5 was in November), peaking especially for long problem-solving sessions on very difficult problems. You can give it a lot more ambitious tasks than what you're used to, the model "gets it" and it will just go, and it's never felt this tempting to stop looking at the code at all (but don't do this in prod!). The model still has quirks that people will run into and the safeguards are configured to be a little too trigger happy for launch, which can hopefully be tuned over time.
I feel a lot of things changing as working software increasingly comes out on a tap. The Jevon's paradox kicks in and I feel my own demand for software growing substantially. You can ask for anything - explainers, visualizers, dashboards, bespoke single-use apps (e.g. a full wandb that is hyper-specific just for your project), you can 10X your test suite, auto-optimize code, run giant research projects with custom HTML for the results, anything! "Free your mind" (Matrix ref). Really looking forward to all the things people build!
Claude Fable 5 is by far the most ridiculous model that makes me genuinely afraid for the future of software engineering.
I compiled the top 10 most unbelievable things I've seen Claude Fable 5 do today:
— Migrate a 50M line codebase from Stripe in a day (humans take 2mos)
— Draw amazing 3D graphics a) Boeing 747 b) space simulations with >5000 objects c) Minecraft roller coasters d) full photorealistic forest scenes e) NYC skyline f) stormy clouds)
— One-shot Pokemon FireRed the game
— Optimize a real world proprietary interaction net evaluator 10x more than the next best model, gpt5.5
AND it's about the same price as GPT 5.5 ($10/M input, $45/M output) vs Fable 5 ($10/M input, $50/M output) and 6x cheaper than GPT 5.5 Pro.
@claudeai Fable 5 built me a Minecraft clone in 37 minutes.
one prompt.
~3K lines of code.
$12 in API costs.
biomes, caves, ores, day/night, mobs all of it.
Minecraft took Notch 2 years.
@claudeai is the real OG !
we are cooked 💀📷 👇
Introducing Claude Fable 5: a Mythos-class model that we’ve made safe for general use.
Its capabilities exceed those of any model we’ve ever made generally available.
Your margin is my opportunity: AI version…
The biggest surprise of 2026 is that the capability gap between the best open-weight/source models and the best closed models has narrowed much faster than the pricing gap. The pricing gap remains enormous while the capability gap is quite narrow.
What does this means in practice?
For a company consuming 1 billion input tokens and 1 billion output tokens per month:
GPT-5.5 Pro: ~$105,000
Claude Opus 4.8: ~$30,000
DeepSeek V4 Pro: ~$5,220
DeepSeek R1: ~$2,740
I asked ChatGPT what it thought about this and it answered as follows:
“If I were building a company today, the economic frontier would look roughly like:
DeepSeek V4 Pro / R1 for high-volume inference.
Claude Opus for premium agent workflows where reliability matters.
GPT-5.5 Pro only for workloads where its incremental capability demonstrably produces enough business value to justify a 20–40× token premium.”
Most CEOs have no idea that, instead of this nuanced approach, their teams are running amok internally by picking the most expensive models in most cases and burning through massive budgets with zero governance, audit ability and control.
As control planes like our Software Factory become more standard, you can expect the run rate revenue growth of the frontier labs to go down meaningfully and the revenues of the open models to skyrocket.
Why? Because we can implement the nuanced approach above and be agnostic to model - instead focusing on customer intent, model task and cost management among other things.