“It may not be feasible of them to say that they’re wrong.” Eli the Computer Guy joins The Tech Report’s Isaac Pound to talk about the uncertain future of frontier AI model development as usage numbers for Anthropic’s Fable 5 have stagnated and OpenAI says they plan to slow down the development of their next model.
ADVERTISEMENT
"Specific models to do specific tasks" I believe that already eixsts. It's called SOFTWARE
For the sake of the world the sooner this all burns down the better.
Lol tokens are way more expensive than employees, one employee burning tokens like crazy has a 2x ouput. While that same $ can pay 5 people using local free LLMs. I’m seeing that daily now.
“We will need 10,000 more data centers to solve the strawberry problem” - Sam Altman probably
As someone in the tech sector, I have not seen anyone in management being realistic, or even understanding, the ROI of AI in their own organizations. I would bet this is pretty common.
Imagine running a restaurant that keeps giving out free food....
all they need is another trillion or two and they'll get it all figured out
Welcome back Isaac! Couple of issues with some of the points raised. LLMs are not cheaper than a person.. If they appear to be then it's because the price is discounted. But how many "actual" jobs has AI replaced? Often the call-centre type jobs which use cheap labour. And you still need staff as backup. The other point is the "usefulness" of AI. Multiple studies have shown that coding with AI is slower. It feels like magic, but overall it's slower. Then there's the dumbing down of people who start to rely on AI. And that is only going to get worse. Then there's the cost. Remember when Openai and Anthropic both had to seriously cut rates? Because it's simply not worth it.
I entered the business world in the early 90s. Back then companies used to upgrade all their desktop PCs every 3-4 years. Dozens, even hundreds, of PCs being swapped out to newer, more powerful machines. Back-end servers were on similar or even more accelerated schedules to keep up with technology. Then 2001 happened and a lot of companies delayed their next round of updates. Low and behold nothing bad happened. The machines they bought in 2097/98 were powerful enough to handle everything they needed. Suddenly businesses slowed their upgrade cycle. I went from getting a new desktop every 3 years to 5/6+ years between desktop machines. The 2008 recession just reinforced this behavior. Companies learned that PCs didn’t need to be cutting edge anymore, they just needed to work. AI and software are much the same. Outside of specific use cases cutting edge doesn’t really add value to regular work where reliability and cost are far more important.
Those OpenAI execs quitting just before being able to afford their own Yachts is a very interesting point.
Is someone going to tell him that the only reason the tokens seem to have a reasonable cost is because open AI and anthropic are setting billions of dollars on fire with subsidies?
As someone who has used everything up to the latest models as of this writing - the attention is the thing that kills it for replacing people in business usage. It's great for getting through a brainstorming session quickly and directing the trajectory of work, but then it starts interfering with the process once it's attention becomes saturated with the context of the actual work. The harness only mitigates this issue to the extent and you don't need to give it less context to know what it's purpose is - and distilling it for a purpose will help it make fewer application specific mistakes, because it's not looking filtering every context down in real-time - it's pre-baked. This goes for every model I've ever used. It's basically a contextual database of prior-art (because, that is what it is trained on). The takeaway is that you still need to people who know what they are doing to get the project across the finish line with any level of stability/polish/safety - AI is basically just a decent research librarian standing in front of the world's dataset.
AI replacing workers is the ceos wet dream of making more money while getting rid of those pesky humans .
If the reliability goes from 85% to 93% you still have to spend the same amount of work to check everything because something is going to be wrong at some point
Sam Altman doesn't "believe" in anything. He is a conman who has himself been conned by his own tech and his own investors. He is a desperate man who feels like his life is over unless he can somehow get the big win. Like any gambling addict, he has no choice but to keep doubling his bet in the hope that he'll win the jackpot and pay off all the debt he took on to make all his previous bets.
Stop the subsidies and see how popular all these AI services are...
Imagine what good could have happened to the world if those trillions of the dollars were spent on improving the lives of people all across the world rather than on ai.
The problem is that all the automation needed in a company can be accomplished by running opens source models in-house. My company is a state-owned cooperative that supplies IT infrastructure to the Dutch education system and academia. We have our own small-ish AI hub running on our cloud infrastructure, and it solves all the needs of us coders and developers and that of researchers and data analysts. I think that the way to go is smaller AI hubs or even local run LLMs.
Eli saying that the current models are doing great and businesses are finding them really useful is hilarious... In a sad way... Credibility lost. Go into any workplace and ask people if ai made their jobs easier or made their work place more efficient. It hasn't.
How can that be? I thought we were curing cancer? Sam Altman is a serial liar and fabricator, the OpenAI board tried to fire him but the ignorant employees demanded he be brought back because he promised them millions of dollars with an IPO, that might not happen either.