J Bex

@Hesamation For me it was ever since the 1m token version of Opus 4.6. I tried to switch back immediately but they removed the original model

AMD Senior AI Director confirms Claude has been nerfed. She analyzed Claude's session logs from Janurary to March: > median thinking dropped from ~2,200 to ~600 chars > API requests went up 80x from Feb to Mar. less thinking and failed attempts meaning more retries, burning more tokens, and spending more on tokens > reads-per-edit dropped from 6.6x → 2.0x. model stops researching code before touching it. > model tried to bail out or ask "should i continue" 173 times in 17 days (0 times before March 8). > self-contradiction in reasoning ("oh wait, actually...") tripled. > conventions like CLAUDE.md get ignored because there's less thinking budget to cross-check edits > 5pm and 7pm PST are the worst hours, late night is significantly better. this means the thinking allocation is most likely GPU-load-sensitive.

@schvik @GaryMarcus That's completely backwards and wrong

@GaryMarcus Hallucinations are not solved...at all. However, they can be reducded, even a prompt can make a huge difference: "You are an expert programmer" -> model confidently hallucinates non existing libraries "You are a novice programmer" -> model does (re)search, uses skills, etc

I have an ML professor asserting that hallucinations have been solved in recent frontier reasoning models. Please supply any data you know that supports or rebuts this claim.

@DalePaulCRAIG1 @IRanMediaco Trump just said they will wipe out an entire civilisation, in those exact words. Who's the fucking maniac?

@IRanMediaco Apparently Iran leaders don't understand the President we don't need your fucking oil or your maniac killer ways were tired of dealing with and you know Trump isn't playing so we may see after 8pm or not up to you

BREAKING: Iran's IRGC Statement: "Self-restraint is over. Our response will now be carried out without any consideration, depriving the U.S. and it's allies of oil and gas for years. If the U.S. Army crosses red lines, our response will exceed the region’s boundaries."

@rllycutegi57931 @EatonEmmerich @Ax983000899227 @10x_er Try a find and replace - or even finding references in an ide - for 500+ occurrences, you will end up with more headaches than if you just used an LLM

@EatonEmmerich @Ax983000899227 @10x_er thats my original point. you can just find and replace and that works for 99.9% of cases. Other .1% can be done via regex. Sure semantic aware replace is better but I would only do that for like Java or something.

Yeah the future of engineering is cooked

@hmziqrs @10x_er The only people cooking you are those who've never worked on a real project with hundreds of files. Your point is valid not only for variables but any pattern you want to change across the codebase

@10x_er BURH! why I'm getting cooked I obviously didn’t mean renaming one variable everywhere. I meant refactoring a module into something more general and now half the function, and variables need to be renamed to match the new purpose. Better let LLM do context aware renaming.

@davidhine99 @aShubhamz Microsoft has always shipped buggy products even when bill gates was there

@aShubhamz dude, bill gates hasnt had anything to do with windows since fkn internet explorer. its that Satay Nuttella cunt thats turned it into the vibecoded sloptrosity that we're currently faced with.

My open challenge to Microsoft Owner Bill Gates: I dare you to click this button and actually stay signed in. This is your product, right? So, come on live and show us, stay signed in and prove it works. Open challenge

@Brezia_Tomson @akafaceUS He was sat with them, probably just picked it up for them

@akafaceUS Some stay behind near tables to pick up purses, wallets. Smooth Moves. :/

A group of people doing CrossFit ran past a restaurant, and diners who saw them assumed there was an emergency. Thinking something was wrong, the customers jumped up and ran away alongside them.

@FecRip17 @AIWarper My random generator had Michael Jackson singing Enrique Iglesias in Spanish, but I think if you translated the lyrics then it wouldn't rhyme

@AIWarper Is there one that can take a Spanish song and change it to English?

I just found the most ridiculously uncensored music generator that I haven’t seen anyone else talking about…. Yall try “Sonauto”.ai yet? Foolish… basically can make any existing artists music 😅 Cooking an example to show here shortly but needed visuals

@PovilasKorop Opus 4.6 does this, it stopped for me yesterday when it hit a wall, turned out it was my fault

I just realized a "bad side" of LLM behavior. So, LLM has a mission to HELP you. Good, right? But. As a part of its mission, it will very rarely (never?) say "I can't do that" or "Sorry, I failed". It will try SOMETHING. Failed? It will try a few other ways. And burn many tokens. In the screenshot, Sonnet is desperately trying to fix a failing test. 20+ minutes. Still running around in circles, but cannot admit that it just can't deliver.

@StayAntsy1 @shahh Everyone saying it's CZ or Trumps fault, but it's just bitcoin doing what bitcoin does

Largest Dip we've had since Trump has Been in Office. My Raw thoughts. 1. Trump has friends that want in cheap, so he gave us a dip. 2. The world knows Trump wanted to be Crypto King so they are selling to make him look like a fool. 3. Trump and his team played us all haha

Can someone explain to me like I’m 10 years old, how Bitcoin has gone down -$32,000 with NO BOUNCE in 14 days?

@TonyJen48667368 @Rajatsoni @Angry_Staffer No he just thinks for himself and tries to get to the bottom of things. For some stuff the right is logically correct, and for other things the left is logically correct

@Rajatsoni @Angry_Staffer He just plays both sides to maintain high audience numbers

Joe Rogan is telling his millions of listeners the Trump admin is acting like Nazis. Shawn Ryan is telling his millions of listeners the White House is actively protecting pedophiles. Meanwhile, the paid GOP shills on Twitter want you to believe Trump is as popular as ever.

@recovery_your Ayahuasca was the only thing that pulled me out of that, I still feel the benefits of a positive and rational mindset 5 years later. It carries risks, but if it's life or death, it's worth a shot. There's retreats you can go to in Portugal or Peru

After 6 months of waiting for mental health therapy. I have been denied it by the nhs but sent an apology. So I have to fight back because if not I will die. Suicide will win. Everyone tells you to reach out get help but there’s nothing to get, no help, no support, just here some medication numb yourself up and fuckoff

@ChombaBupe I had an algorithm in a private repo that I worked on for 3 months, there was no public solution to the problem. Now chatgpt can solve it with the same solution that I came up with. Coincidence? Probably not

I deleted my GitHub account a long time ago because I was suspicious that if I hosted my work in a private repo on GitHub Microsoft would still access the code. From the looks of it, the hunch was correct because private repos were scraped for training Copilot.

@floridaman11112 @GatsbyHuxley @Soul_Eater_43 @MichaelButtonX Nope, actually anything outside our atmosphere could stay there even for millions of years

@JayBex_ @GatsbyHuxley @Soul_Eater_43 @MichaelButtonX Wrong

Still can't wrap my head around this Modern human brain-size: 1,000,000 years old Dawn of civilization? Only 6000 years ago Why did we live as hunter-gatherers for 990,000 years? And what changed?

@rune2059739 @Lorenzo_Del_T @HamEggsnSam @MichaelButtonX It's actually much much smaller than how Atlantis was described

@Lorenzo_Del_T @HamEggsnSam @MichaelButtonX Atlantis was a civilization from before the last 12k years. Richard Structure fits with the description and there is water erosion seen from space around the area of it to the sea.

@floridaman11112 @GatsbyHuxley @Soul_Eater_43 @MichaelButtonX Satellites could stay orbiting for 1000s of years, not decades

@GatsbyHuxley @Soul_Eater_43 @MichaelButtonX Without any reboosts, the ISS naturally would deorbit within a couple of years. To keep it in orbit, we have to do reboosts on it almost monthly. Our furthest orbital satellites would take a few decades to work their way back.

@che1i0s @burkov It does, but they do quantize the models aswell. There should be regulation to make them disclose if you are using a quantized model

@burkov Depends on context length probably, what about prompt templates/agents md? When you constraint it just triggers different part of terminators brain and can lead to a better outcome. Treat it as multivariable surface you are approximating

I was just about to post that as well! I'm sure Anthropic cheats and serves a weaker model when overloaded or needs GPUs for something urgent. The last two days, I feel like I'm talking to a retard from 2024.

@Dupuisa1 @RickKoeppen @WomanDefiner She had her wheel full locked to the right, away from the officer, that's the key fact you are missing. The officer shot through the side window, she was in no way using it as a deadly weapon

@JayBex_ @RickKoeppen @WomanDefiner i'm canadian and it would be the same here

Because no one has done it I lined up both videos time wise, put them side by side so you can see where the officer was when the first shot was fired in both views. He was in the process of being hit by her car.

@Dupuisa1 @RickKoeppen @WomanDefiner Classic American logic..

@RickKoeppen @WomanDefiner Regular resisting arrest does not warrant deadly force you're right. But since she was in a car, she was using a deadly weapon to resist. This does warrant deadly force.

@swingerinsure @JerisGibbs @Inevitablewest Venezuela is in South America lol

@JerisGibbs @Inevitablewest Last I checked, Venezuela is in America, so it is our business…

This was how Maduro's men attacked peaceful Venezuelan protesters. This is what the left is defending.