that’s the point. These things are for getting the first 80-90% in relatively good shape very quickly, and then humanity takes over and finishes the job.
Just to give people an idea of the CAPEX levels going into physical AI Infra.
The way I see it, if they cannot monetise AI by next year…we will see a correction.
Ran into a new (for me) AI error. Was playing around with using AI for getting some investment advice for a particular scenario. It was reasonably good at suggesting appropriate funds for what I wanted to do. Where it screwed up is the “perfect” fund it suggested for me was discontinued in 2013 and the next best alternative it suggested is closing in June.
While AI may “know” everything, it seems to have massive gaps in what it recalls. I’m not sure how solvable this is. You’d have to go down a lot of rabbit holes on each prompt to prevent this issue which I’m thinking would push up compute costs substantially.
Many young folks are wisely seeking careers in the trades.
AI adoption is currently male-skewed (based on UK research) but the roles most exposed to AI automation are dominated by women.
From an economic standpoint, the usual argument here is for this segment of the labour force to upskill themselves and/or change careers but the speed at which this is all happening makes this very difficult.
I think that because of this we are very likely to see markedly higher structural unemployment in the short to medium term until new jobs are created on the back of AI understanding/skills.
I’m predicting that Oracle will be the big domino to fall and they’ll get a bailout because so many companies rely on their databases and systems.
Larry Ellison being a buddy of Trump will just be a coincidence when Oracle gets bailed out.
2 Likes
A recent experience with ChatGPT 5.5 Pro | Gowers’s Weblog
A recent experience with ChatGPT 5.5 Pro
We are all having to keep revising upwards our assessments of the mathematical capabilities of large language models. I have just made a fairly large revision as a result of ChatGPT 5.5 Pro, to which I am fortunate to have been given access, producing a piece of PhD-level research in an hour or so, with no serious mathematical input from me.
I use AI for my proposals now, it rights 100% of them.
Mind you, I read every single word and confirm every single number, every single time. And it requires a few iterations, normally slashing sections. But otherwise, it’s fine.
I also use it for website content because it can actually write stuff a bit better than me - it does a decent job of motivating readers, calls to action, benefits vs features, stuff like that that I get a bit sloppy on. Again though, I read an approve every word.
Um…hopefully you proofread those proposals? 
5 Likes
I treat content from AI as malicious. 99% of the time it will do a great job to lull you into a false sense of security. Then every 10 proposals or so it will sneak something into the middle that’s completely absurd. Like a negative number or something.
I see AI being a good use of AI for research paper’s abstract. What I’d be interested to see is what % of the actual body of the paper is AI-generated and/or AI-influenced.
1 Like
Mythos still not released, though Open AI has since dropped 5.5 which also crushed the AISI hacking test, though didn’t do anything else special.
Anthropic has also gone from a nobody to a big player overnight, and as a result has been desperate for more compute, buying more from Google, Amazon, and space x
Anyway, more shots from the department of disturbing graphs:
I would think a LLM writing assistant would be very helpful for writing science articles, especially since many scientists are not native english speakers. And my sense is that generally the insight of a scientific article is not particularly linked to the wording of the article as compared to some other fields like history, english literature, etc.
2 Likes
I think that the wording is very important as it is conveying the author(s) interpretation of the results along with laying out the arguments of how the results/observations are supporting the interpretation/conclusion. Neither of these items are not very well suited to AI, IMO.
I agree that the wording must be very precise at certain places. For example, it is important to say xyz is a 95% CI interval calculated as such and such, etc. For those parts, even the non native english speakers probably know that technical language very well. Or there is are essential elements of the description of the experimental apparatus that must be just so. But those essential elements are connected by other more mundane words, and the LLM can probably help with those.
As I think about it more, I realize I was thinking more of a shorter articles that just report a result.
I agree that a longer article with a less predefined organization that contains a lengthy argument would probably be less helped by a writing assistant.
IMO, you write in your native language first, then look to translate it as necessary.
I’m pretty sure that AI would be good for that use . . . just do a "translate from native language to other language . . . the do a translate from other language back to native language . . . then compare back to the original as a first level check.
But there should also be a rigorous peer review process that can also help clean up translation quirks.
1 Like
If i’m not mistaken, LLM’s are really really good at different languages. From what I read, they don’t view language as ‘english’ or ‘french’ distinctly. I don’t really understand, but i think they’re a step back from there being distinct languages.
1 Like