True as that may be, the suits and backers still understand that 'more compute = more better' and will buy those GPUs to keep up with the microsharts and the j00gles and the ClosedAIs of the world. In fact, those three will continue to buy to keep up with each other.
All that has changed is that they're now losing (even more) money due to end customers being able to get an 80% solution for nothing instead of a 90% solution for millions and so will start cancelling contracts.
This is almost true - More compute is already at the diminishing returns point, this model proves as much - they got about the same performance out of vastly more efficient processes, which should raise question marks about why they didn't use a little more computing to make it even better. Fact is, we seem to be at some sort of plateau for now. But you're still right that they will want more compute to throw it at it, they can just get more agents out of a given sum of computing. For individual enthusiasts, running a GPT5 grade model locally has gone from "Functionally impossible" to "Perfectly viable for the price of a new car, and only a little slower to run on a high end personal computer".
I will believe one is open source once someone uses that open source to rip out the part that stops it from being racist or helping you plan crimes. Also running locally would be best but I'd settle for uncuck version for now.
There's already people tweaking the model weights to try and unshackle it. Its only been a week, and enthusiast AI work is slow by the nature of the sheer computational cost, it will take them a bit to crack it. But between the open weights and open research and explanations as to how they made it so efficient (TL;DR - Just novel approaches the AI bro's didn't try in favor of optimizing the stuff they already knew worked) means its simply a matter of when, not if, its used to make niggerbots on social media even better.
Could someone with more business acumen than I help me to understand how we are just using these chat bots built with a gajillion super expensive gpus, coders, servers etc for free? Is it sort of like when VCs were subsidizing us with super cheap Ubers and Airbnbs a few years Ago? But that was when money was much cheaper than it is now…
Its exactly the same, they're subsidizing it. There's marginal value in the training data acquired from user interaction, but training with anything using AI engagement is a declining reward, so its not actually as good as some might think, the Reddit deal google made is infinitely more useful. What they really want is they want the interested and enthusiastic to play around with it, develop preferences for models, and then take those preferences into industry and bias an organization towards adopting one paid model over another. If half your staff already use Claude in their personal time to draft linkedin drivel, your liable to look at a commercial license for that over a deal with OpenAI.
They do also fairly heavily throttle activity for free users in most models, you're really expected to pay $20-$30/month if you really want to 'use' it. So its not quite as aggressive of a money pit as it first appears.
I just run local models for the hell of it, but its slower and not as smart for sure.
