I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives. More thoughts below.
How I do code reviews these days: more code = more selective human review i.e. don't read all the code.
1. Every PR gets a multi-agent first pass. It should find bugs, verify them, rank by severity, suggest fixes. Approval stays a human call on anything that matters.
2. Low blast radius changes on less sensitive code (there's often a lot!) can skip a deep human review once that review is clean. This helps keep the explosion of PRs manageable.
3. Core / sensitive paths still need an owner and human sign-off. That's where you spend time: verification, constraints and earning trust in what the agents can safely cover. You want to keep recoverability.
Agents do the first pass and humans cover blast radius.
Great question from @GergelyOrosz!
There are 30(!!) upcoming Omarchy Meetups around the world in September now! Absolutely gorgeous artwork for them. Signup in a city near you or start your own. Transfer the enthusiasm! https://t.co/ZbFktXHDq5
There are 30(!!) upcoming Omarchy Meetups around the world in September now! Absolutely gorgeous artwork for them. Signup in a city near you or start your own. Transfer the enthusiasm! https://t.co/ZbFktXHDq5
No one is ready.
No one is ready for cancer. No one is ready to get laid off. No one is ready to lose a child. No one is ready to start a business. No one is ready to settle down. No one is ready to lose weight. No one is ready to get divorced. No one is ready to say goodbye to their parents. No one is ready for unexpected depression. No one is ready to see their best friend commit suicide.
No one is ready. No one is ready. No one is ready.
Stick this on your bathroom mirror. Tattoo it to your leg. Never forget.
Ready isn't a feeling. It's a decision.
You must violently decide to lean into the uncertainty and take action. Being action-orientated is a behavior and 99% of people don't do it. As soon as you even think you're ready, the chaos of the world reminds you nothing is predictable.
But there's a hidden strategy. A counterintuitive truth that makes taking action easier. Once you start taking action your fears get smaller because you now have new information about what works and what doesn't. So there's less to be afraid of.
Make big moves. Make the world violently respond to your deepest desires. Make mediocrity your enemy. Make action replace overthinking.
No one feels ready — not me, not you. And definitely not your heroes.
You'll never feel ready. Just start.
Every org is building an agentic SDLC in some capacity, but retrieving the right code is super tricky. Here are three big issues you will run into.
The first issue is that similarity is not relevance.
Ask an agent to fix a login bug, and retrieval (grep or indexed) will surface every file that mentions login: auth routes, session handling, middleware, and even a DB migration. All related, but none of these explain the actual failure.
The file that matters is usually reached in subsequent steps and by following the evidence, and not by retrieval ranking against the first sentence of the request.
The second issue is that capacity is not attention.
A bigger context window does not mean better understanding. There is the classic 'Lost in the Middle' problem with large contexts.
Codebases are full of "convincing distractions" - dead implementations, generated files, test doubles, parallel code paths. Every extra file that gets retrieved, especially if it is dead, can negatively impact inference.
The third issue is gullibility.
If we give our agent several plausible files upfront, it can build a confident theory around them, make changes in the wrong files, and form an incorrect understanding of the codebase.
By the way, these issues will always be there. You will have to navigate around them and build your agentic SDLC. Most of the solutions will be specific to how your org codes.
Hope this helps.
Schrödinger's code
1. fable wrote it, it must be great. in fact, i had Sol, opus 5, and Grok all review it.
2. if i look at it, its horrible
state changes based on observation
Looking and behaving better amplifies your ability to get anything in life. Applies to tons of engineers. Spent last week in India where these problems are especially acute.
I put together an opinionated list of 15 things everyone ought to do:
Fitness
– Not only is being fat terrible for your health but it’s a big signal that you don’t take yourself seriously. Diet is the single biggest contributor to being fat or skinny fat and takes 6mos to change your body shape. Most things around you is carbs-slop. You are very likely eating way more calories than your maintenance. Start eating 1-2 meals a day, no snacks. High protein (1.6g/kg/day). For a 70kg person that’s 112g protein. Read the labels on food. Remove sugar, that’s a carb. Remove fried food. Oil is extremely calorie dense: treat it like the plague. Gets easier after 2 weeks. Don’t let parents peer pressure you into thinking this isn’t healthy. Don’t let friends peer pressure you into eating “just this one meal”. In India, there's a 100 myths about how protein has daal or ghee is good for you or a little sugar in chai doesn’t matter. “It's just a biscuit what’s the big deal?” when they add up to 500 cals a day in carbs. Or "our grandparents ate this and didn't need protein, this is just a fad." Take GLP-1s / Ozempic if it's right for you.
– Work out 3-5 days a week. Lift heavy weights. More muscle makes you look better but also increases your metabolic rate. Do an hour every morning first thing. Use weights first, focus on cardio later. Lift as heavy as you can, you won’t look jacked (both men and women). Prefer compound exercises: bench press, squat, deadlifts. For cardio, get to running 5k. If you hate running, incline walk for 30mins. Cardio can’t beat diet for weight loss. Consistency beats intensity. If you can, get a treadmill in your house.
– If you smoke and drink, I know it’s hard to change habits overnight and I don’t think it’s easy to quit cold turkey. Just try to pick the few days you’re going to do this and try to moderate on the order days to make those days special.
– Get 7+hrs of sleep if you can.
Grooming
– Do your hair. Don’t do a fancy haircut. Keep it short and simple. Cut your hair every 2-3 weeks. Put some product (matte hair wax) after a shower and blow dry your hair into shape. No, it doesn’t make you bald faster. Use potent anti dandruff shampoo regularly if you flake. Use conditioner.
– If you have glasses, get LASIK if you qualify for it. You will not lose your eyes and you’ll look way better. If you must keep glasses, find a shape that works for your face shape.
– Shave everyday. A clean look shows people you put in effort every day. Some people can pull of the beard look. 90% of people think they can but look terrible. If you are doing a beard, trim and shape every day.
– If you are balding, get a hair transplant. It hurts but usually being bald hurts more. if you have obvious balding, just bite the bullet and go bald.
– Clothes. Make sure you have the common colors for all your basics (white black navy grey and some shades of beige burgundy). Don’t wear things with logos. Wear fitted clothes. Dress for your age. Don’t wear jeans. Don’t wear plain thin T-shirts. Find colors that suit your skin. Overdress > underdress. Have tailored formals. Shoes should be stylish but minimal.
– Accessories. I’m more opinionated on this but I think clean is best. Never forget your shades when it’s sunny. Usually wear a belt. Maybe a watch (I personally find smart watches very tacky). If you need a backpack, keep it minimal but super high quality. Ideally, no tattoos. Ideally, no chains, no rings. Never do too much: nothing beautiful asks for attention.
– Teeth. Get the dentist to scale your teeth annually and whiten them. Use veneers if you need them. Use Invisalign to move your teeth in shape if you have bent teeth. Use mouth strips for bad breath. Use a tongue cleaner. Floss at least once a week.
– Scent. Wear some. Not cheap stuff. Try to carry it on you if it’s hot and you sweat. Shower every day. Use deodorant. If you sweat a lot, make backup plans for clothes.
– Skincare. Face wash and moisturizer every day morning and night at minimum. Use sunscreen every morning.
Manners
– Smile more when you meet people! It’s the easiest hack to make friends and be more social.
– Make sure you always monitor how much you speak vs the other person and moderate yourself. Talk about yourself less, and the other person more. Actually listen to the other person. Let your work speak for itself. If the person doesn’t want to talk to you, don’t force it.
– Speak better and speak slowly. Improve your clarity and your diction.
– Be authentic, honest and reliable. Relationships build over the long term. People remember fake, social-climby people who put on fake interests to fit in. Remember people’s names. Be on time.
Doing all of this helps open a ton of doors professionally, helps with the dating, helps make friends, keeps you healthier and improves your self confidence. People take you seriously. Of course, there are plenty of exceptions.
At the end of the day, all of things is polish for your main main thing: the work you do or family you build. Anyone who makes fitness or style their main thing ends up not being liked by anybody (unless it’s literally their job). Many people do not realize that these things also help contribute and open doors in the real world.
I see so many people asking: why is Ramp building their own coding agent? Spotify an agentic development environment? Thomson Reuters their own AI model? Why do they BUILD and not BUY these tools?
It's because AI made building cheaper, and more "profitable." If you want your engineers to become "AI engineers", they need to build stuff.
So why not build tools that are better than what already exists? With much less effort than before? As a side effect, getting "cracked" with AI infra, and buildign with AI?
Not following the AGENTS.md standard and sticking with CLAUDE.md is becoming a problem at scale
I haven’t felt it much since I mostly use Claude, but the friction is obvious when moving workflows or agents between tools
Standardizing such a critical file, one that’s sent with every LLM request, helps the whole ecosystem move forward
Let’s see how this plays out
I’m thinking about banning Claude code at Shopify until they change their mind and read AGENTS.md and .agents/skills etc.
Insisting on only reading CLAUDE.md sometimes leads to split brain problems when different team members use different tools. Just unnecessary.
File systems and object storage, that's all I am going to read and write about from now on. There is an incredible amount of untapped potential in this space. Weekend project incoming.