Der leitende KI-Direktor von AMD ist der Meinung, dass „Claude sich zurückgebildet hat“ und dass man „man ihm nicht zutrauen kann, komplexe Entwicklungen durchzuführen“.
https://www.pcgamer.com/software/ai/amds-senior-director-of-ai-thinks-claude-has-regressed-and-that-it-cannot-be-trusted-to-perform-complex-engineering/
19 Kommentare
>Claude […] cannot be trusted […]
Yeah buddy, i don’t need to be a director of AI to have understood that months ago.
The latest is more combative and less productive. Which is really annoying when you try and inject the correct solution to thinking Opus’es as they burn compute arguing two incorrect interpretations of the code.
Its almost like this half baked tech is getting shoved down everyones throats and it dosen’t do what is claims to.
The plagarism machine cant be trusted to run on its own, its a limited tool with limited applications that needs to be used by experienced developers.
The current state of autocorrect alone has convinced me that current gen „AI“ implementation is a huge mistake.
The [issue report](https://github.com/anthropics/claude-code/issues/42796) from Stellar Laurenzo, Senior Director of AI at AMD
A lot of people think they sandbag the old model’s performance right before they release a new model. So, the new model looks better than it actually is.
With them hyping up mythos now, it would make sense that they start making opus worse.
Pre-Mythos tuning
Thats because the best engineering is behind company firewalls. AI has never had an original thought, its just an amalgamation of the most statistically common answers from humans. It cant create a new, novel approach. At best it can recognize statistical patterns
Like trusting the kid who always cheated on his homework to come up with their own thoughts.
lol if the amd guy is saying that out loud, it’s probably worse than people want to admit. feels like everyone’s just pretending the demo version is the real thing
Their technology can’t keep up with demand. Subscriptions exploded while Sam Altman literally bought up 1/3 of the world’s RAM supply and everyone else is scrambling and salvaging parts. Add to the fact that data center space is at a premium with all the energy it takes to power and cool the GPUs, and it’s a hot mess. Anthropic decided to deal with this issue by reducing the amount of compute everyone gets instead of reducing the amount of users who access said compute.
Source: Am guy scrambling for parts and data center space for AI.
“Maybe now the bubble will finally burst and all these companies will go bankrupt” when will you all understand. AI isn’t for us. They need AI, and in order to be able to work on it openly they had to come up with “reasons” that we need it or should use it. It’s completely useless to us and they know it. Its only purpose is to eventually control us. They have literally told us this to our faces.
Of course it’s regressed, it had a big influx of users.
„trust“
Only C-level folks use that and AI in the same sentence. I work with the stuff every day. Trust, as in blind trust, has never been something I have even considered. The tools are good – fast, useful, etc – but at the end of the day they are just tools. Trust comes from the engineers that review and approve the changes.
I’ve had less issues and more reliability with Claude than I’ve ever had with OpenAI
>complex engineering
This is generally where most AI / LLM / coding tools have fallen apart in the real world. For those who aren’t close to developers, code is most often a series of steps to get to „done“. It’s akin to getting ready in the morning:
– Turn off alarm
– Shower
– Get dressed
– Leave for work
AI is fantastic at simple tasks like turning off your alarm and showering. It gets flummoxed when it comes to more multi step processes and may put on your shoes ***before*** your pants.
But where it really falls apart is in connecting between steps. *Couldn’t get your pants on? Close enough, let’s get to work!*
Source: Not a dev, but I pretend work.
Can you really ever trust an ai 100%, probably not! But yeah, for complex engineering and many other tasks you really need to be able to trust it
But that’s the secret Captain, AI has always been untrustworthy.
I wonder if this is somehow related to AMDs lack of involvement in Claudes new „Project Glasswing“. All the other big players are involved, I’d be curious to know if AMD has a specific reason (beyond what this guy says) for that.
Wasn’t Claude just the best last week and now it is like the worst?