This challenges the long-term necessity of proprietary frontier models
One possibly unpopular opinion I have is that peak model intelligence will be irrelevant for most real world use cases, and within a fairly short period of time sufficient intelligence will diffuse very widely through open source.
Positive users agree peak model intelligence is already irrelevant since smaller local models suffice for most tasks, while negative users argue frontier advances remain essential to cut failure rates.
No Digg Deeper questions have been answered for this story yet.
@xlr8harder I can see it already. >70% of enterprises’ token usage is to extract data, save it in a saas crm and draft emails
Ironically peak intelligence is necessary to fix the bugs introduced by AI because these bugs are too complex for the people in charge to fix without SWE background
@xlr8harder this is where apple may assert their impact. local, 100% private GPT5.5 level on-device models could divert a LOT of demand from the big labs
@xlr8harder solving CRUD slop will be enough for almost everyone
@xlr8harder When do you predict this will be true?
@xlr8harder I think this is wrong because models have relatively high failure rates for even basic tasks, and failures are expensive. You'll pay for models that are capable of vastly more than your task requires to minimize fuckups.
@foomagemindset Economically useful tasks.
@xlr8harder sufficient for what?
@xlr8harder Improving ocr results (by far my most used solution at my workplace) worked on gpt 3.5 and now on a 7b mistral. Added bonus is that it was one of the only apps that didn't go down after they killed 4o.
@xlr8harder People have no idea how boring most million dollar ai savings look
@AndreTI That's why the op is referring to the future.
@xlr8harder Right, but currently there's no way to raise the floor without making the model more generally capable across the board.
@xlr8harder Peaks and troughs are strongly correlated. If you want shallow troughs you also need high peaks.
@AndreTI But as the capability floor raises, many tasks become trivial and you don't really care if your model can solve millennium prize problems if you're using it to handle customer support.
@advancedjd CFOs will be going into shock over AI budgets this year and people are going to be motivated to find cost savings very soon.
@xlr8harder Model intelligence is already sufficient for majority of real world use cases
@xlr8harder yeah staring at an excel spreadsheet is just too easy for the models
@xlr8harder True. My main personal agent runs GPT-5.5 but I’ve moved many simple jobs to Qwen 3.6 35B running on an AMD AI MAX.
It’s more than capable and essentially free.
@xlr8harder Well peak human intelligence is irrelevant to most real world use cases too!
@xlr8harder It's already true. Token cost is the bottleneck for so many agents right now. Just need cheap tokens, discrete tasks with guardrails and you can do so much.
@xlr8harder kinda agree cause most real world use cases run on supply chain and operations which is hard to replicate, also open souce models will have really good shot at being almost best as frontier closed source