GPT-5.6 just made itself CHEAPER

ChatGPT says the Max setting is a waste. Use high.

You really must prioritize talking about Chinese open source model EFFICIENCY. The cost gains per task efficiency out does Americans models and many many of your subbs will gravitate towards these Chinese models. Js saying

Improved itself? They probably used the same method Deepseek did and distilled the larger model.

I just hope competition does not result in safety minded actors being acquired by megacorps.

Competition can not just be about price and potency of a product if we care about safety and other concerns.

Fable to Pro and 100% next?

I want to see Fable 5 price drop!

I agree. I’m surprised more people aren’t talking about this. It feels like the beginning of real RSI, and if it delivers, it could give American AI labs a meaningful commercial edge over Chinese labs. If competitors don’t catch up quickly, this advantage could last for quite a while.

You assume price == ecffeciency. Im guessing luna is under used and scam altman needed to distract you.

Makes sense. Distill from your own models.

need to be comparable to your own pc to prevent open source to clame top spot.

That's why we should keep Chine on the loop of everything. They will keep the leeches from sucking too much blood

Is Luna basically like haiku though? also does this mean I will also be able to get more out of my codex usage?

Recursive self improvement is now necessary to remain competitive...

💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩💩

Yeah I've already decided to stay with OpenAI Chatgpt for now... The way they did this make it more efficient made it more accessible. That is a lot better than the elitist exclusivity of 'Anthropic' and Dario's arrogant statements saying we are now in the era of 'expensive tokens'.. I know, intelligence is a private good but considering he made money out of the collective knowledge of he internet, its unfair for Dario to make this access highly 'inaccessible' by making claude use soo expensive!

The only reason OpenAI and Anthropic are ahead is because the US prohibits sales of advanced chips to China, and China doesn't have EUV lithography yet.

Once China even gets halfway to catching up in EUV and starts making remotely on-par chips, it's game over for OpenAI and Anthropic. But that might not be for several more years.

The usage on the subscription didn't improve tho

luna and terra is terrible. Charts are just bs.

Musk might be able to reach them because he might throw more compute. He has largely caught up, if he can continue have ore compute, more spend (and he can) and doesn't do too badly in the talent department... The other one that cold is Google, but they are currently a long way behind and would have to make a lot of use of the massive compute they have.

This isn't recursive. Models are trained from data, and the algorithm hasn't changed. They aren't just a code to optimize. These are just task running platform upgrades or hosting upgrades. Escape velocity is only hit once it can optimize and reduce the training part, and not even but find a way to keep improving without more data, which is a real-world physical limit. Not the runtime.

When comparing the work of Luna Terra with models even such as Opus 4.6, you will be surprised that on paper they should work better but in reality they suck.

so more slop for cheap

Waiting for china to do some "efficiency improvements" and 1 up the price drop

this is why we need regulation

Anthropic should use fabel to do the same and become at least a little efficient.

wtf it's free.!

you’re running open claw 😂 and you’re nervous about open ai making improvements??? 😂😂😂😂

It's hilarious how everyone counts OpenAI out, done, finished, prepping for bankruptcy, etc.

THEN BOOM.

5:40 I thought they are not supposed to train on our data?

I have been running codex 20hrs per day for the last 10 days on Sol Extra High, Im only on the 100$ plan

How many sponsors per video, lol. I stopped watching

Its worth noting the graph looks like logarithm on x axis - not linear. Does not change your argument - but worth noting

China is paying you. NBC is now pushing media narratives of "America needs an open source solution" - that would be cancer to persue harder than it is

Necesitamos que habilites la traducción a español.

How are they going to pay back the VCs?

this is 'recursive self improvement" with diminishing returns. Next iteration, it'll find far less places to improve, but faster, and then even faster to discover even less.
This is not the singularity. Until it can start coming up with actual breakthroughs, it's just finding errors fast. And, no, programs that you've found errors in don't just magically have more errors.

4:51 they wrote up a blog post .... "Sol wrote up a blog post" 🤣

4:30 after using Sol the past few months I would say that not many OpenAI engineers know what that means. Sol is great at doing stuff which is way over people's heads and making up a legit, engineer sounding, name for it. 😆🤣

… anyone else remember where we were at in January?

Can’t imagine where we’ll be by 2027, the rate of progress is mind boggling.

I don't get it. Even the best frontier models screw up so often. Why would I want to save some money using a cheaper model that will certainly screw up even more and add to my manual error checking and verification work? 😅

How can I trust that they are giving us unquantized model APIs, i randomly see that the model quality goes nose dive and then gets back up at random hours. i'm unable to prove or quantify it though.

Looks like someone’s trying to maneuver themselves out of bad press