I'm just surprised he didn't rename it to American Intelligence.
30,562 karma · joined December 4, 2018
That's literally what happened. Thousands of .si domains were bought up before his renaming, and then offered up for millions.
PS: His wife is from Slovenia. I'm sure that's a coincidence...
The "expensive part" of generating the next token is streaming in the model weights from memory. The computations are relatively simple, which is called a "low arithmetic intensity" in industry jargon.
So what they do is batch multiple chats together and compute the neuron activations for all of them together.
This is vaguely similar to how some database engines work, where if multiple users need to run a "whole table scan" query, the additional users "join" the streaming workload of the first query mid-way, then loop back around to complete the first part that they missed. The AI accelerators don't do this looping, but the concept is the same: amortize the expensive I/O over multiple computations running in parallel.
The "turbo mode" token rate thing is almost certainly your query getting sent to slower or faster hardware, like B200 vs newer B300 kit.
My approach would be per-user (or per-project) "memory".
I don't mean a tack-on like a RAG / document store with "notes" added to it in the background, but an actual medium- and long-term memory. Something like an extension of the KV cache stored in High Bandwidth Flash (HBF), or a subset of the weights trained "online", similar to LoRA.
This would not be transferable to any other base model, so would be excellent "lock in".
The downside is that the memories likely wouldn't be transferable to new models either, but I can imagine solutions to that too. I.e.: Train an MLP to "translate" from the old memory weight space to the new one.
Azure doesn't exist.
Where the humans are the pets and the (machine) Minds are the real protagonists, in control of everything?
Where the machines have crafted a language that affects the way people think, aligning them better with what the machines wanted?
That's your utopia?
I do hope you don't work for any frontier AI lab!
That’s the root of the problem.
Even if the data centre raises billions in revenue and pays billions in taxes, that isn’t paid to any government entity close to where the building is.
The locals get replaced by machines, the revenue is pocketed far away.
It’s the equivalent of big mining companies “investing” into mine automation in some developing country. The locals do not benefit!
My firsthand knowledge of these is that yes, they happened.
However in most cases OpenAI informed the victim, not the other way round.
Even medium to large organisations are typically unable to detect a breach unless it does obvious damage.
Everyone can notice their whole network getting encrypted for ransom.
Few can notice SQL injection blended in with application requests that contain SQL snippets in the normal case.
From what I’ve seen in the logs, the “AI agent breaches” are almost polite for the want of a better word. Like a gentlemanly catburgler picking a lock to sneak in and… take a picture of a rare artwork in a private collection.
Ass covering instead of responsibility.
Security theatre, in other words.
When the consequences for failure are very high but personal reward for success is very low, everyone does everything they can to avoid being held responsible for the consequences.
Note that I didn’t write “avoid consequences”!
That’s different.
Roughly:
Get API keys for multiple vendors or just use OpenRouter (but availability of frontier models tends to be limited). Alternatively, Azure Foundry has everything except Google models, so just two subscriptions is enough.
Run the same prompt and same input image through each of your chosen models.
Then feed the smartest model the original image together with the collected output texts. Use a prompt along the lines of "Merge these attempts to OCR together into an corrected and improved combined version, taking special care to exactly preserve the original's typos, etc, etc..."
You can do this manually, it's just fiddly. It's not hard to automate, most of the "code" is English instructions!
The downside of this approach is the cost: even the "light" frontier models are a few cents per page, which is not so bad until you're doing this 5x or 10x times per page and suddenly scanning a notebook can set you back tens of dollars, more than buying a good novel at a book store.
I picked up on this technique back when GPT 4 was released. People noticed that it could translate ancient Akkadian, but only if you ran the prompt through 4x times and merged. I tried this with a few random samples I found online and the merged translations were generally better than the "official" ones, even thought the individual attempts were unreadable gibberish.
There are already scripts/tools floating around for this!
Look into OpenRouter Fusion, Consensus AI, Multi-Model Debate, etc... or just whip up something yourself.
GPT 6.1 and Gemini Flash 3.8 both do pretty well, their OCR of your sample image is only "wrong" in the sense that the original has typos and they corrected some inadvertently and/or filled in gaps where you had "unintelligible" in the canonical text.
If you have the budget and want the best possible results, you need to run each image through multiple models and then combine the outputs into a final "merge these" prompt. Better scanning helps too, your sample image is rotated and you used a phone in low light. Try a DSLR or a flatbed scanner and process only one page at a time instead of two at once.
GPT 4.5 was scrapped for similar reasons.
You misspelled “exfiltrated” for the purposes of AI training and NSA dossier building.
It turned out that all of our peers supported blob storage better than SFTP, which has some “show stopper” problems like forced outages caused by mandatory host key rotations.
I've noticed that in "programming" forums, questions have devolved.
In the "before times", you'd get a mix of simple-medium-hard-extreme questions.
Google and StackOverflow made most of the medium and hard questions disappear.
Now with LLMs, all the higher difficulty levels except "beyond extreme" have vanished, leaving only the stupidly difficult questions that nobody can answer -- and hence doesn't contribute to collective knowledge, including the AI's knowledge.
Conversely, there is a new level of "below simple" questions by people that I dare say are too stupid to even attempt to become software developers. The type of questions that can could always have been answered with just Google, and are "beneath" both AIs and humans. These are worthless questions along the lines of "my computer not work, plz halp!?" followed by "I can't reboot it!" as if that makes any sense.
Again, that contributes nothing of value to the zeitgeist.
Worse, because the effort required to spin up a blog is trivial now thanks to AI and people are getting desperate for work, I'm seeing a mountain of "slop coding blogs" that are either 100% low-cost-AI output full of errors, or organic human-origin slop... full of errors. No longer are blogs written by the "best of the best" to disseminate their hard earned knowledge, they're now an alternative to a resume and produced in bulk by the most junior developers hoping to land a job through "exposure".
This also massively dilutes the last remaining source of good public information to draw. Google searches are useless now because finds nothing but endless SEO garbage.
What's going to happen in 5 or 10 years from now, when all of the pre-AI knowledge[1] is stale and out of date, making it irrelevant, but there's no new knowledge being publicly produced to train on?
Will programming languages, tools, and frameworks ossify because new languages are "post-AI" and hence untrainable due to a lack of StackOverflow questions and answers?
Will AI labs build their own private "AI members only" version of StackOverflow, building up their internal knowledgebases for training, sharing this with nobody else?
Will humans be paid to solve problems for the sole purpose of generating AI training materials? Is this the future of knowledge work!?
See: https://github.com/dotnet/SqlClient/issues/767#issuecomment-...
"So, let's say you created a new pooled connection and closed it, driver will pool this connection but the physical connection is already active on server side. When the driver reactivates this pooled connection, it's only done on client side by taking it out of pool stack and setting it's state Active. But when next query is executed, we ensure a "RESET" bit is set in TDS header with next command on this newly reactivated pooled connection to cleanup any stale state, as per standard behavior."
Oracle requires an opt-in(!) flag that is incomplete and configured on the server instead of the client, which is a "special" kind of nonsense that I have no patience for: https://docs.oracle.com/en/database/oracle/oracle-database/2...
Non-determinism is injected, on purpose.
I used to obsessively check the assembly (machine code) output of my C++ compiler because it would regularly miss essential optimisations and my tight inner loops would blow through my time budget.
I haven't done this in decades, it's just not worth the hassle almost always, the compiler will do a better job than me at optimising the runtime performance. If anything, "second guessing" the compiler will often degrade performance!
I'm starting to notice this now, where my clumsy attempts at refactoring AI-authored code introduced bugs that the original vibe-coded output didn't have.
This can be compared to protectionist policies such as tariffs.
If employees are hard to fire, then companies hesitate to hire them.
If employees on-shore are protected, then un-protected employees overseas are more competitive, eventually gutting local industries.
Etc.
This came up in Australia recently where a liberal government proposed loosening some employee protection laws with the idea that it would encourage hiring.
It did. It worked.
Then the policy was swiftly reversed by a more left-wing party, and a lot of those newly employed people lost their jobs because they turned from an asset to a liability.
Protectionism is a double-edged sword.
Coloreds are still hugely overrepresented in the criminal "justice" system, they're still much poorer on average, are still disenfranchised en-masse through gerrymandering and voter ID requirements that accept shooters' club membership cards but not ID cards poorer blacks might possess, etc.
For all intents and purposes, the USA has a caste system that has evolved from an earlier system that was apartheid in all but name.
There's no "meeting of minds" and there's no exchange of value if I simply benchmark a piece of software and then don't use it any further.
This is a bit of a grey zone legally, and steps all over a bunch of consumer protection laws, free speech laws, etc... depending on the jurisdiction.
IMHO governments should explicitly ban these "gag clauses" because it prevents free market competition if customers are not allowed to benchmark software and publish the results.
Got a reference for that?