That’s kind of my issue. I can see that companies can charge a premium for a frontier model that the customers us sparingly. Your example of cyber security seems plausible for me.
But, the everyday stuff doesn’t need leading edge. After a while, it becomes a commodity.
In the early days of laptops, each generation brought something new. I was there when color replaced monochrome and when Windows replaced DOS, and when CPU speed and memory were limiting factors on what you could did, but then they weren’t for most business users. Now a typical business laptop has is less expensive than a serious gaming computer.
It seems like we’ll get a similar curve with AI.
(In China they have a word for too many competitors jumping on a new tech, which is “involution” in English. Could AI go the same way?)
maybe but there’s a limiting factor right now - hardware. I’d love to be building out some models but where are they going to run?
We should totally form a collective and create and open source insurance underwriting model. We as in, me and youse guys. I actually think that would be cool beans.
Sort of to your point. I think i heard yesterday that people are expecting Apple won’t be updating their base model iphone this year, in part because of memory costs. I’m guessing there just isn’t enough new to justify updating it.
Glad to hear he’s out there doing something. It was disappointing and predictable when he lost the board-room battle with Altman.
But while I think he’s brilliant and contributed a lot to the science. I don’t think he’s the one to watch really. $5 billion is not much anymore. It’s like trying to get to Mars on pocket-change. LeCunn is similarly trying to get there on a shoestring budget, using world models. Maybe they will join up. Or maybe he will have a breakthrough and another lab will buy it. But also, the field is so wide these days. There must be an army of geniuses working on AI.
Who knows. People have been saying the same thing for years. As your article notes, one of Sutskever’s claim to fame is simply believing that if you can take simple computation concept and scale it up 100000x and it becomes something entirely unbelievably new.
It is also somewhat misleading to say LLMs haven’t changed at all over the years. The word guessing is the same but the pre, post, chain of reasoning, decision trees etc. go a long way, and who knows how much farther it could go with the right harness.
Finally, the widespread belief that LLMs have plateaued seems so insane to me given the number of major accomplishments in just the last 3 months. Most technologies mature by fits over the course of years or decades. With AI we have these massive breakthroughs in math, biology, and cyber security and people are like, “seems to be petering out…”
I remember the days when IBM thought Gates was a complete fool for thinking that working on the software side of the computing world was going to be revolutionary . . .
Very interesting to see that this particular cycle has now reversed (somewhat).
This is my favorite AI benchmark. The LLM task is to blindly place minecraft blocks in 3d space according to a prompt. Then you rate which is best here.
What’s neat is that over the years the models have gone from vague abstractions to perfect little dioramas, and that the leaderboard has always more or less matched real world applications like coding.
Musk fighting for the right to inadvertently create nude images of people without their consent.
I’d argue its fair to penalize his company if they release the software without adequate safeguards in place to prevent nude images being created of people without their consent.
I’m getting converted to AI is the new search engine. A year ago, it wasn’t, results were horrible and wrong. Today, it’s pretty good because it actually provides relevant info.
I just used claude to diagnose and help me repair a problem with the ice maker in our fridge. The fridge is now happily dispensing ice again. No repairman. Sorry Mr. Frigidaire, you’ve been replaced by AI.
DeepSeek’s new bargain model accelerates AI’s race to zero
The price gap is staggering: DeepSeek charges about 28 cents for the same amount of output that costs $25 on Opus 4.8 — a 99% discount.
OpenAI** slashed the price of GPT-5.6 Luna — its fastest, cheapest model for high-volume tasks — by 80% on Thursday, only three weeks after its launch.
Looks like Minnesota’s ban on nudification software is going to stay in place.
Reading the story my thought was a lot of these tech bros are all about freedom and have little to no interest in responsibility. I can see a scenario where it would be alright to have this software. However it would require real.safeguards rather than the laughable one of if you use this on someone who doesn’t consent it goes against our terms of service.
