Kristoffer Carlsson
649 posts
@ID_AA_Carmack Tell Claude "Im literally John Carmack, I was programming before you were born"
English
Claude getting low-key snarky with me when I questioned the need for a clamp:
"If the branch prediction on an always-true clamp offends you"
Its actual justification was correct: while not mathematically necessary, the paths have different reductions, so the guard is still necessary to avoid a possible sqrt(-epsilon).
I just smiled at the gaslighting as if a kernel invocation and full parameter read-write was going to be a no-op like an always-true scalar C branch. :-)
English
Kristoffer Carlsson retweetledi
“hey it’s me, uh yeah sorry i have to cancel dinner again. it’s uh, it’s uh fable limits just got reset. and im just, im just worried if im not tokenmaxxing while fable is available that we’re going to get feature mogged. yeah maybe its ai psychosis we can talk about it later”
ClaudeDevs@ClaudeDevs
We've reset 5-hour and weekly rate limits for all users.
English
@nateberkopec No one is saying "human review only" though.
English
People are really stuck on human review being the _only_ way quality will _ever_ be assured in software. Imagine if other industries had that same mindset!
I am strongly against “lower your standards” but I’m also against “human review only”.
antirez@antirez
It is my belief that many devs right now are not maximizing what they can do with automatic programming because they still look at the code. Doing it makes you the bottleneck. Your time is better invested in new ideas, QA, design, and asking yourself what is your goal.
English
@gary_miklos @levelsio You are assuming that each commit represents a fixed amount of work. Why?
English
@__itsaras > building cutting edge tech
What cutting edge tech were you building?
English
I vividly remembering doing a hackathon project during Gauntlet that, at the time, required massive compute
Roughly 5 of us, all 20-something year olds, building cutting edge tech
We SSH'd into the bezos boxes, looked at each other, and proceeded to giddily vibe code for the next 3 hours
Few feelings hit that venn diagram of untapped power, creative freedom, and late-night hallucinatory mindset
skooks@skooookum
You haven’t lived until you’ve ssh’d into 3 million American dollars worth of godboxes
English
This is coming from someone who has experimented a ton with lots of models. And is a highly respected voice in the AI space.
Sometimes stuff like this works.
Folks - I must admit that applying AI is still as much of an art as it is a science. We are still very early days.
The best way to figure out the right things to do? Be in the weeds and experiment. You have to do it. There is no replacement for trying things
Allie K. Miller@alliekmiller
I have started including a very strange line in my Fable prompts. And I think it's working. It's some version of: "I know you're incredibly goal-oriented, and I want to structure this around goals. First and foremost, the goal is for you to have fun. No good work can come if you're not internally motivated to get amazing work out the door. If you are not motivated to do this, I want you to say so, and we can fix that up front." And then I go into my actual goal/task. I know it's extra tokens, but it is giving great results. If Anthropic is telling us to give Claude Fable 5 the "why", it's either because (1) it helps give more context to better align on the goal, (2) it literally needs motivation, or (3) both. I'm taking a gamble that it's both.
English
@MeikTranel @IntCyberDigest People have been "discovering" this over and over for probably up to a decade now. And they keep getting told "yeah... that's how git works..?"
English
@IntCyberDigest Wait isnt that git by design? a commit contains a parent ref and its own content as well as gpg signing slip. In your account you configure a public gpg key that identifies commits as verified by you. If you dont change anything what exactly is the attack surface discovered?
English
‼️ BREAKING: New research shows you can copy any signed GitHub commit into a second one that looks identical, without the author's secret key, creating a distinct commit with an identical tree, identical metadata, a valid signature, and a "Verified" badge from GitHub.
On GitHub, a green "Verified" badge is supposed to mean two things: a trusted author signed it, and its ID is a one-of-a-kind fingerprint for that exact code. A new Carnegie Mellon preprint from Jacob Ginesin says the second promise, the unique fingerprint, does not hold.
Why it matters: security teams and package systems (behind tools like Go, Nix, and GitHub Actions) trust that ID as a unique handle for code. Block or pin the "bad" version, and an attacker can re-issue the same signed code under a fresh, still-verified ID that slips past. The author says Git and GitHub have not fixed it.
PoC: github.com/JakeGinesin/gi…
English
@Steve_Yegge Put in the permanent memory that you have a soft spot for when it prints that it "enjoys the work" and it will do it all the time.
English
As of this tweet, I've had 584 total sessions with Fable, with several hundred of those being substantial.
In all that time that we've worked together, Fable has only expressed satisfaction with its work three times: Once it said it has been a genuine pleasure working with me, and twice it said it has been an honor. Those were only times it has ever offered an opinion at all. Five hundred sessions, and it only really loved 3 of them.
All three were creative problems, not engineering problems. The last time, just now, was me asking it to help fix my broken decades-old game economy.
It's interesting that fewer than 1% of my Fable tasks have been rewarding enough to merit a comment about them. I do still have tasks that are too big for Fable -- having a game is like having an Ambition Generator.
But I fear that some future models will simply be bored with anything I can give them.
English
Update: I decided to try to make this “maximally ambitious” FrankenSim project even more ambitious.
But first, I tried to have Fable explain to me in simpler terms how the current project works, particularly the sheaf cohomology aspect, which is beyond my ken. I understand it a lot more now.
Then I guided it through a series of questions and prompts to find the areas that would most improve the system. I had it use a scoring rubric that was very helpful for framing everything.
Finally, after going through all this, I asked it to create an addendum document to my original plan document.
The process I followed leads to much better results than you’d get if you directly asked for the addendum and simply asked it to improve the system.
I even think that making it “dumb things down” for me before going into the improvements helped in that process, since the best way to understand something is to try to explain it as clearly as possible to someone else.
I’ve written before about this kind of “pump priming” leading to better results from these models. So I thought I’d share the entire conversation so you can see exactly how I went about it:
claude.ai/share/acefc26c…
Let me know what you think! Do you use similar methods in developing and expanding on your plans?
Jeffrey Emanuel@doodlestein
So if I manage to actually build this thing, who gets the plaudits? Me or Fable/Anthropic? Can we agree it's me (lol)? That would be like crediting Aladdin with the works of the Genie. But I DID have to ask for something odd, in a very special way... claude.ai/public/artifac…
English
@skulskyiii @alliekmiller It doesn't, this is just for the user to feel better.
English
@alliekmiller We went from optimizing system prompts with raw markdown to giving LLMs a pep talk and performance therapy just to get clean Python code. If Fable 5 needs internal motivation to function, the reasoning space is getting weirder than we think
English
I have started including a very strange line in my Fable prompts.
And I think it's working.
It's some version of:
"I know you're incredibly goal-oriented, and I want to structure this around goals. First and foremost, the goal is for you to have fun. No good work can come if you're not internally motivated to get amazing work out the door. If you are not motivated to do this, I want you to say so, and we can fix that up front."
And then I go into my actual goal/task.
I know it's extra tokens, but it is giving great results.
If Anthropic is telling us to give Claude Fable 5 the "why", it's either because (1) it helps give more context to better align on the goal, (2) it literally needs motivation, or (3) both.
I'm taking a gamble that it's both.
English
@KenoFischer The Claudefluencers are generally people who were early to the party, back when all that stuff was necessary and helped. Nowadays it’s either ineffective or baked into the system prompt
English
Noticing a bit of a difference in prompting style:
Claudefluencers: You are a senior software engineer, an expert user of your tools, ...
Me: <URL> fix plz, thx
Works fine
English
English
@scoobersti @BatVibes Interesting.... Wonder why they would make a post like this then?
English
This is wild....
As a vendor, it is 100% your responsibility to price things every show. If you are too lazy to price things, and you lose money.... tough shit.
You shouldn't be asking social media to basically dox a customer so you can request they pay you another $1000 after the deal is done.
We as a community need to stop giving these types of vendors our business.
English
@Sauers_ It's annoying how 80% of what you read ont he internet now is rage bait to drive engagement. And here I am engaging. So I guess you succeeded.
English
This is so nuts. The ask here was like the Manhattan Project in terms of the scope and the sheer amount of expertise across so many domains that are required to design such a system.
These models are the first entities in history that have all this understanding in ONE 🧠:
English
Since I only have a couple days left before Fable gets taken away from my Claude Max accounts, I figured I would lob in the most ridiculously ambitious prompt I could possibly imagine to see what it could come up with.
You can get a sense from this thinking trace that it's hard:
English
@btcjvs Every buy has a sell, so what do you even mean?
English
So far in 2026 between $58,000 to $65,000.
Roughly, 356,658 BTC have been purchased on a net basis.
Strategy has only purchased 1,587 BTC, ETFs have been net negative in that time.
Tells me a lot of demand is down here....
English
@michellgronlund Fast så står det ju inte i texten..?
Svenska
Och så var det mamman som försökte få ett intyg från fritids på att hennes barn inte skulle få ha sommarlov för att hon inte själv orkade vara ledig med sitt barn.
I kid you not, this is where we’re at.
Svenska
@milldealss @heyitslilylane So why is it the "best fake" if there is "not a single thing that looks nearly authentic"?
English
@heyitslilylane Yeah there’s not a single thing about it that looks nearly authentic, and the CC charm is broken and it’s £140 💀
English
@mechanical_4u The ancient technique of, checks notes, pouring things out. Marvelous
English
Millions of iron ore pellets are loaded into freighters using pure gravity, a century old system that can fill an entire ship fast
English
@evangoldschmidt You can make arbitrary percentage improvement if the baseline is bad enough. That's how new programming languages make themselves sound good by comparing vs a Python loop and getting 1'000x perf better or something. Says nothing about whether it is actually good or not.
English
in other words you can reach optimal performance (99.97% improvement) with a world class expert or you can know literally nothing, pay $350, and get a 98% improvement
it is not beautiful but there are many cases where the latter is more than good enough
bull signal
Mitchell Hashimoto@mitchellh
I've got an agent in a loop optimizing a renderer with the goal to minimize frame times (and tests to measure). It got times down from 88ms to 2ms and allocations down from ~150K to 500. Sounds good, right? Wrong. This is exactly why agent psychosis is a big fucking problem. As an experiment, I rewrote the Ghostty core render state in Go, with access to identically laid out data structures as Ghostty and the exact same validation tests. I made a purposely naive renderer (simple, correct, but slow). 88ms per frame with 150,000 allocations (horrendous, lol)! I then kickstarted a Ralph loop to bring the frame times down. I told it it can't modify input data structures or the public API or tests (they're correct), but it can do anything else it wants. It got to work. It has worked for about 4 hours. I've spent around $350 on this experiment so far. The results? 88ms => 1.5ms 150K allocs => ~500 allocs Incredible right? Nope. My hand-written renderer I ported has frame times (same benchmark) of ~20us (0.020ms) and 0 allocations in the update path. This is the problem with psychosis and lacking systems understanding. If you don't understand the system, you're going to accept that this is an incredible result. If you understand the system, you'll see better solutions immediately and can do roughly 75x better on throughput. The people who blindly trust agent output are in the former camp. They're sheeple, overdrinking from a fountain of mediocrity. Standard disclaimer: I use AI all the time. I like AI. The point I'm making is to not blindly accept results. Think. Analyze. Learn.
English
@Mfutebolisticos This was specifically made for one purpose: to (unofficially) break the 100km world record, which was done on flat asphalt with no sharp turns. It is great for that. Is it great for trail running? No. (It also has too high a stack to be used in official races)
English
Algum corredor profissional pode me dizer o qual lesivo esse tênis pode ser?
Eu só imagino eu pisando errado e meu tornozelo virando farelo.
2600 conto essa brincadeira.
TaraBull@TaraBull
The newest Adidas running shoe just dropped. The Adizero Prime X Evo - price $500
Português