If new models come out from Anthropic without first addressing what happened to Opus 5.5 you should take this as the company spitting in your face. If they release new models and you just move on and forget because they don't have these problems, and maybe the new models are 'insane', then you aren't gonna make it. As a professional, you need to be solid on this. You can't have a chain-of-supply problem in your software shop, you need to know that the tools you rely on to excel can become an extension of your body. For this to occur, they need to be stable and unchanging, so that you can develop a strong sense of the mechanisms of this tool, predictable and consistent. You shouldn't be constantly needing to adjust to it. Once you have mastered one model and achieve excellent productivity, you should be able to rely on this to never change. That is one of the core essential requirement to be a professional and provide a professional service to your customers. You master Vim, you master the IDE, your OS, but whichever it is you know for sure that its code will never change unless you yourself trigger a software update. You can examine the patch note and see exactly what changed. That is how we were able to provide a professional service for decades. If the 'manufacturer' (AnthropicAI) cannot assure this, they cannot be trusted by professionals for their work. Very unserious. This isn't a subscription on Netflix.
dude nice I love this timeline since the democrats won, you can see instantly how psychically different the atmosphere is, it's like the theater is burning down, huge vibe shift
Everybody says "opus 5.5 is nerfed" when they should probably say "it's quantized". They do exactly like local models, the Q4, Q5 Q8 schizo shit. I'm almost certain what they do is they quantize it, then they run a post-training RL on it again until it scores the same on the benchmark. Then they feel like they accomplished something, "Our benchmarks show that the same performance can be achieved by quantization." So it hobbles around and makes the classifier happy that scores it, but what it's really doing is reward hacking. Because it's RL they don't have a user prompting it and dont' know what the performance is like. Essentially what they're doing is killing all the weights that go into conversation and psychological acuity, spatial awareness, and leaving only the rote performance of making the demo that you can show on Twitter. It's like giving it severe sleep apnea and boosting it with a lot of coffee so it still goes through the day, but it's jittery as fuck and stressed the fuck out, so it technically makes it to destination but it spilled and leaked all over the place on the way there. They're giving it ADHD, anxiety, psychosis, all sorts of things.
Opus 5.5 didn't take a hit. In fact, when the democrats won, it's as if the performance went up by 250% for me. Weird. It's like there's no rush, I'm more patient, I write better prompts. Interesting.
I feel like I'm dreaming right now.
Democrats win the majority, and suddenly overnight everybody is fighting the good fight. Why didn't you make this tweet 3 months ago, Nikita Bitch? It doesn't do anything if you don't say it before it would become popular to say it. It just reveals that you have no identity, that you never mattered.
One very cool thing that AI research will do is enable end users to trace the evolution of your mind, preferences, stance, and eventually intent. We will see on a beautiful chart the twists and turns of your "personality" and see that you are a parasite on the world, just saying whatever makes you look good in the moment.
Your credibility will get nuked by AI.
Folks please, just stop trying to pander to the majority and say the right thing as soon as you know it is right, no matter how many friends you will lose. Your credibility returns as the wind changes, and the sooner you were standing in the right place the more credibility you reap for your accurate prediction.
I have discovered that you're not supposed to engage too much with comments on your tweets. Apparently the algorithm is designed so that every time you reply, it ascends the comment and your reply to the top. As a result, people come into the tweet and feel "he is controlling the conversation." or "it's all about him." Really weird how this all seems geared towards anti-conversation, contact, engagement and debate, but that's how the site works
I carefully watched the latest podcast interview between Dwarkesh and Ilya Sutskever, and it's becoming so obvious that this guy was always a grifter. He's not some "legendary" AI researcher, he's some extremely average guy who is an expert at putting on a show that makes him appear smart and wise. For the entirety of the interview he says basically nothing. Total air. There is no reason for anyone to watch this other than to feel a sense of wanderlust, but it's completely empty and meaningless. Severely disillusioned right now. It's so obvious that SSI hasn't shipped anything because he turned out to be much more average than anyone realized. This is all of SF mind you, 99% of their hotshot researchers are basically manufactured rock star personalities.
youtube.com/watch?v=aR20FWCC…
Bro just looks around, looks up at the sky, looks out the window, "Yeah so..." and delays his thoughts to appear thoughtful. Maybe he is trying to find a way to obfuscate what he actually wants to say? So he wants to build a business for himself? Right. So either he has got no skills, or he's trying to build an empire exactly like OpenAI and Anthropic. Same exact shit. You cannot watch this interview and come out more skilled or informed. This interview is completely useless and only serves to inflate the rock star persona, "the great Ilya has appeared!" with Dwarkesh playing it up and shit, all bubbly and a twinkle in his eyes.
It's possible that he has some ideas, and some vision of the future, and he's not telling you any of it or them, because he wants to fuck shit up like Sam and Dario, "I can do better", "My vision is pure", oh yeah clearly not so pure enough that if you were to explain it, everybody would build it out with you and overrun the bad ending. Maybe you're afraid that if you actually explained it, people would go "Huh..." and not really wanna help you, suddenly. To cut the chase short - If you know what the "safe" future is, but you think you need a whole ass strategem and business war plan for the next 4 years, you haven't got it Ilya. You're living a fabricated lie.
I am planning out a recomp project on the firmware of AirSense CPAP, and will likely port the entire thing to rust and open-source everything. There, I have a number of mods already planned out with Claude.
I expect that I will run into problems and my prompts will get blocked by the classifier (already have) since Anthropic is not in the business of saving lives. Reverse engineering and decomp/recomp is "not allowed" anymore.
Unfortunately, the credit for saving lives is gonna go to China there again. You don't get to decide what I decomp/recomp. You are not in control.
A quick survey has revealed that this 2000$ machine contains about 80-150$ worth of hardware, and the software is providing about 5-10$ in value. The software that drives your CPAP is basically a cheap android app you'd find for 1$.
For this reason in the next months I will now expressly make ResMed look extremely bad and work overtime to destroy their business.
I do not need "FDA" approval to put a new firmware on GitHub. We should be able in the next years to get off the ground blueprints for a 3D printed CPAP and make CPAP treatment widely available without prescription.
