In a recent interview on The Joe Rogan Experience, Marc Andreessen stated that the very latest frontier models have finally enabled artificial general intelligence (AGI), which he describes as AI that’s on par with humans.
I also saw this, in a response from Gemini:
You have run into an active, confirmed bug in the Gemini API.
Right now (as of early June 2026), there is an ongoing regression specifically affecting the Gemini 3 and 3.5 Flash endpoints. Google’s backend executes the web search perfectly, but it silently drops or empties the groundingMetadata object from the API response payload before it reaches your server.
If we actually had AGI, wouldn’t Google have used it to fix this bug? Theory: a lot of what we hear about LLMs being as good at coding as people, may be hype.
Right now we have word and pixel regurgitators. Yes, the output is very convincing, but what its actually doing underneath the hood isn’t very close to any kind of actual intelligence IMO.
I highly recommend Cal Newport’s podcast episodes titled “AI Reality Check”. As a computer scientist involved in the digital ethics field, I find his perspective very insightful.
Meta (and all the others) seems to disagree. Meta is building the largest data center ever, campus, a 4,000 acre campus that will use 5 to 7 Gigawatts of power (as much as New Orleans on a peak day).
So interesting! Especially with regard to Facebook – the only U.S. company with the ability to spend enough to create frontier models plus the interest, at least so far, in releasing LLM models open-weight.
Well, if the LLMs are going to keep improving, I guess it makes sense to continue creating bigger ones. OTOH, it kind of seems like in a couple of years a lot of companies will run open-weight models on local computers to lower the cost and reduce the outflow of corporate data.
I guess it’s a bet that the AI industry is making – the closed-source models will keep improving so substantially that a huge volume of spend will go to them. And if that bet is the winner, it’s worth the billions going into it. A company wouldn’t want to get that bet wrong if it has the resources to be a player in this space.
Another big part of it is that many are expecting AGI to pop up magically out of a hat as the models get bigger. My opinion at this time is that, that isn’t going to happen based on current technology, which just predicts the next word based on work in the materials it was trained on — work that was created by – as Steve Wozniak puts it – AI as in actual intelligence, i.e. humans.
Another explanation for the data-center spending - use cases like this are going to be too big for companies to host on their own hardware. So companies will still need these data centers, and they’ll want to work with the market-leading companies, e.g. openAI, Anthropic, etc., even if they don’t need the frontier models. Plus, more use cases like this are bound to come up over time.
