This is really insightful. Based on this he should speak more. We do not understand these models. Emergent behaviour or generalization appears in the simplest of models. Such as the Grid world tasks. The model learns to solve many goal cells just on the basis of the limited training it receives. To study how these models evolve we should go back to first principles and look at the smallest and simplest units to understand what is going on.
Look let's see what we know. It definitely makes us as individuals more productive: we write more, I derstand more, can even code when we don't really know , ' hello world'. The missing link is the macro. In economics theory millions of optimised utility functions add up to a macro environment. Now in case of AI it seems the macro is not clear. I think we need to distinguish between enterprise AI - yet to evolve - and individual AI - current AI.
It's likely thst the principles of economics apply to the llm companies as well. In textbook economics, monopoly leads to a host of efficiency problems: they charge more to consumers and innovate less, just because they don't need to. Ultimately everything costs money. A classic example of monopolies hijacking development was India in the 1970s when we followed the import substitution policies leading to large public and private monopolies. It lead to a slow stagnant economy and very low innovation. Things speeded up in India when we opened up. Similarly an AI monopoly is a bad idea. Right now the big labs form a oligopolistic structure. However it is the fear of opensource developments which keeps them investing and innovating.
This ignores the fact that frontier AI companies have to attend least break even. To do that they need scale. If they limit to the US market, there is definitely going to be a slow down in innovation. R&D is a costly exercise especially at Frontier model scale. So, restriction to markets is bad for these companies. The US may be a large market, but the world is a big place.
Yes, the US govt's regulatory actions are also going to push lots of people towards chinese models. Corporates outside US are going to look long and. Hard at their processes and identify which ones to outsource to a Chinese model and which to keep for Claude/gpt. It doesn't look good. Might have been better economically to have kept quiet and handled problems. But transparency is good. It will lead to better regulation and lower profits for the big guys. Other companies and open source will come up.
Gpt 5.5 has improved a lot, especially in coding. Still has to improve on sycophancy. It tends to mostly agree with what is suggested and then proceeds to execute the code. Some kind d if suggestions, improvement or even at times challenges might be good. Ofcourse the big defining edge if gpt is the personalization. In my opinion that is the no. 1 driver if stickiness. I remember that everytime I open a fresh chat with claude and start with the story so far...
Agreed. Denis hassabis definitely. He looks as though he still does 'researchy' stuff. The other 2 definitely strategize and talk at high levels and take updates from their employees. But, I could be wrong g. Maybe their statements about AI are deliberately vague, do that you don't get to know what the lab is doing. At least when Denis talks, be used AI concepts and terms. Sam and Dario both thrive on generalusms and. Vagueness
I wonder if they even have the capability to regulate? Ai is something everyone talks of at a high level. But to regulate you need to have almost as much knowledge as the lab that created it to test it thoroughly. Has the US govt got such a dept or staff? How does this regulation work in practice
@pmddomingos Ai just takes away the barrier to learning. But the fundamental reps of reading, understanding and remembering, anakyzing, etc still have to be done. Each step is easier, which means we can go faster but that comes at a cost, as we retain much less.
You know I wonder about these 'stars'. Being Indian we know that most people are expendable and replaceable. At least that is the way the Indian corporate sector operates. I wonder about these AI stars...how are they do irreplaceable? How have they generated such strong intellectual monopolies.
@karlmehta Satyas idea is a good one. But there has to be a large mothership model which is then finetuned with corporate domain knowledge. But to be useful the model needs not just a company's data, it needs to know about competition, related products, the supply chain, etc. it's complex.
@karlmehta I think we have always shared the planet with other beings. The intelligence levels of most animals is severely under rated. Now we will also have digital beings as well. Till robots evolve into cylons, I think we are quite safe from the singularity - but with competition from AI
You know these model rankings and hence the model company ranking keep fluctuating. Everytime a big release happens - they become buzz if the moment. However, I would not write off any off the major models whether it be chatgpt or gemini. They have their strengths. For example chatbots personalization is incredible and leads to a lot of stickiness from customers. Anthropic on the other hand have gone for a more efficiency led strategy where you really have to be conscious about token burn and limitations. There is a tradeoff: lots of people don't want the stress if monitoring their tokens.
The Indian IT companies thought they had a permanent moat based on cheap labour arbritage. Unfortunately the ground has been pulled from underneath them. Nobody is saying that they should have become an Openai, but having a R&D program where you build some inhouse models? That's not too big and ask. All companies do that. At least they would have people who understand the end to end process. Indian IT is characterized by acute short term myopia institutes by project based billing cycles. Being in R&D is probably a misnomer for being on the bench.
@TVMohandasPai That is is so nice and optimistic and non critical. Unfortunately, lack of debate and criticism is probably what holds us back from change. It's so easy to say, ' look where we came from.. '
@LocasaleLab The most important next horizon to be conquered is domain. There are no easy answers here for either AI engineers or AI models. You have to ask the right question or seed the right thought - without that I doubt even fable can do much.
@burkov It was a very poorly directed movie. Very long. The beginning was too slow for a movie- people were getting really restless about where is this going. The middle with car chase sequences was ki d of watchable and the last third was like, 'when is this going to end'.