Zuck doesn’t care, but it sucks to have worked in that group. The VR stuff was stupid without reference to AI. I don’t blame him for not foreseeing that AI would be big, though Meta seems to be screwing AI up also.
I think Zuck read too many sf stories where VR is indistinguishable from reality, and thought he could do that also. I know someone who works at Meta, in hardware, and she organized a session at a conference a while back around this, and it seemed very culty to me.“VR is the future, VR is the future.”
I typed out this whole long rant about Meta’s VR strategy, before realizing we were in an AI thread 
In that case… all I can say is, I hope AI RAM demand settles down in the near future (or production increases) so the Steam Frame can actually see the light of day. It’s okay if Skynet takes a few months longer!
It will be interesting to see which of the many AI strategies are going to go bust at the cost of billions. Progress only gets made by exploring many false paths as well as successful ones, but these false paths have very high tolls.
Incidentally one of the best uses of AI, especially ChatGPT is as a language tutor and conversational partner. One reason is that AI can carry sophisticated conversation on practically any subject in many languages. If you want to improve your language skills especially in a particular area, say landscape painting in French or automotive technology in German there is a no better tool for doing it.
I’ve been messing around with Claude. I thought to ask it to write an extension for Brave Browser.
I’ve been looking for an extension that would allow me to select some text in the browser, right-click it, and select from the context menu a putative item, Search Apple Music, and then it would automatically, well, search Apple Music. Haven’t found one. So I asked Claude to make me one.
And it did. Works perfectly. The only thing I didn’t like was that it opened the Apple Music website instead of the desktop app. So I asked Claude to have it use the desktop app, opening it if not running. It did that. Pretty cool!
Recently, I’ve been using Google Gemini to help me build a taste profile of myself. I’ve had it ask me questions about what I like to eat, and so far it’s come up with 17 of what it calls “New Frontier” recipes that I might consider trying.
I’ve also been using it for entertainment purposes. One conversation involves me playing as Blitzø from the Helluva Boss TV series and going on missions in the multiverse. My favorite thing was when Blitzø sat courtside at a basketball game next to Barack Obama when an attempt was made on Obama’s life. Blitzø stopped it and traced its source back to Secretary of Defense Pete Hegseth, so he captured him and took him to The Hague.
That triggered a response from Donald Trump in which he threatened to invade The Netherlands, so Blitzø enlisted the assistance of IMP ally Batman to literally scare the shit out of Trump and make him change his mind.
I also am using it to track my progress as the Kent State Golden Flashes in College Football 26 for the PS5.
There was an impactful RAM optimization that rattled memory stocks a couple of weeks ago. Google published a technique to reduce the RAM needed for the LLM context cache (not the model itself).
There are a lot of articles on TurboQuant and it’s effect on the market, but this article has a more measured take:
The bigger win will be reducing model size without reducing accuracy, but as the article points out that might just increase demand.
And demand, of course, is going to be a function of the amount of investment capital dedicated to AI firms.
So for those of us who would like a RAM upgrade sometime in the near future, the crash can’t come fast enough…
I crunched some numbers and while making little films in AI is really cheap (compared to what I imagine it would cost to actually film or animate the equivalent using traditional methods), it’s not free.
A 5 minute video (generated by Kling.ai 5-15 seconds at a time) would cost $500 to $1000 or so, depending on your plan. And that’s just generating the video in a single take. That doesn’t include time and costs on multiple takes for each clip (which is guaranteed) and for creating reference images, writing prompts, actually writing a story, etc etc.
Although I imagine you could recreate Avatar for a lot less than $200milllion+. The N’avi would just have six thumbs is all.
I had an interesting one today. I was listening to German songs and didn’t understand a few words in one of them, so I asked ChatGPT for a side-by-side translation (note to @EinsteinsHund: I quite like “Serotonin” by Isolation Berlin).
The LLM explained that it couldn’t publish the full song because of copyright law, but it could break the song into chunks and then do what I wanted. So, it gave me an artificial split of parts 1-5, each translated.
Also, one bit of weirdness: the third time it listed the chorus, it wrote “Part 5 (Chorus повтор)”, so the Russian word for ‘repeat’ dropped in. Strange.
Well … when you speak every language in the world, there’s bound to be some leakage between them. 
It sure happens with humans.
I have found Google’s AI extremely helpful to provide strategy tips for video games.
I think this is a perfect use case: It’s a subject about which there is a lot of information on the internet, and much of it may not be accurate, but a summation of the consensus opinion is likely to contain a lot of useful information. It’s a subject about which nobody has any financial or ideological incentive to deliberately spread false information. And lastly, it’s a case where it doesn’t really matter if it’s wrong.
It occurs to me that another point along those lines is that the information you seek is found in very diffuse form.
i.e. there’s a hundred sites and 500 articles / posts each of which contains some small piece of the overall puzzle. Making it impractically laborious for a human to use conventional search then follow and read each link to assemble the gestalt.
Yes! The random Russian word dropped in a big wall of normal English happened to me last week. And I got another one like that with Georgian. I asked if why and it just said it sometimes mixes words up or something like that. I asked why Russian and it said it wasn’t targeted at me, just arbitrary.
I just used ChatGPT for the first time. It’s good at these things:
I really enjoy having philosophical discussions with AI. ChatGPT has been my AI of choice for the last few years and it’s been good but I decided to try out the paid version of Claude and , man, I have been really impressed with Opus 4.6 which seems a cut above in terms of writing and thinking quality. Had a fairly long discussion about AI sentience and its implications and I was really impressed with Claude the way it pushed back and probed the weaknesses of my argument.
Personally I am excited about the prospect of becoming an amateur filmmaker in my old age. I reckon in 10 years or so it will be possible to use AI to create coherent narratives with persistent characters and without today’s visual glitches. Moore’s law and general algorithmic improvements should greatly lower the costs too; I suspect a feature-length film for a few hundred dollars of compute should be quite feasible by then. It’s going to be amazing.
I think it’s actually possible now. A lot of tools let you create “characters” by uploading a few pictures from a few different angles so you don’t get “character drift”.
I think there’s a bit of an art to the prompting as well. Too simple the model can veer off in whatever direction. Too complex, I think it gets confused…and veers off in whatever direction.
So like much of AI, it’s like working with a sort inconsistently competent crew.
It’s sort of the same thing at work. We’re implementing a bunch of AI systems to automate some insurance stuff. But a lot of it is still sort of wonky.
Really the biggest proponents of the “AI gonna take yer jobs!” narrative seems to be people who sell AI and career coaches.
Yeah my working assumption is that the cyborg era where humans and machines will have to work together will last fairly long perhaps 15 years or more before you get true AGI where AI can just do advanced tasks easily by itself. One reason is that real life invariably throws up a near infinity of edge cases and weird anomalies that are likely to trip up AI.
I think the relatively slow roll out of full self-driving provides a nice reference point. Driving is a nice intermediate problem, much more complex and open-ended than say Chess or Go but obviously much less complex than huge domains like business and all of life itself.
Eventually though I do think that AGI will be solved and that will certainly be a profound phase shift for all of humanity…
While I don’t disagree with the idea that AI code can contain errors in logic, as distinct from coding errors, those stem from the same kinds of assumptions and deficits in requirements specifications that affect human software designers, too. In particular I reject the notion that “An experienced programmer has learned how to anticipate all the cases” which is why thorough and systematic testing methodologies are so crucial.
One of the books that was influential in my career in software design was “Software Reliability: Principles and Practices” by Glenford J Myers. It was filled with many insights on the achievement of reliable, structured software design throughout the development lifecycle. One of the chapters that discussed failures of an early US missile defense system was titled “When is the moon not an incoming missile?”. The problem here is self-evident, and it was “experienced programmers” and software designers that made the blunder, not AI, in creating software that implicitly assumed that a rising moon was a threat to national security.