AI Reading for Sunday May 19
Thinking about the demos last week, I’m kind of skeptical Astra type assistants will be practical soon.
Processing video on these models is around $5 per hour on current costs.
I don’t think it works quite as well as these models suggest. I’ve uploaded video and the model sort of gets a lot of what’s happening but can’t answer detailed questions, I think it needs to get trained for specific tasks.
The privacy implications are considerable for a model that semantically understands everything you are doing and seeing, loads your home or workplace to the cloud,
Probably will work a lot better and get cheaper though, especially if you can connect to a local server appliance, for instance in an operating room where an assistant can be a surgeon’s second set of eyes.
I think the hype and one-upmanship dynamic and competitive accelerationism can get a little out of control and people can release or tease things that aren’t ready and where they haven’t thought through all the implications. Self-driving for instance might be OK in a Disney parking lot or Phoenix but it is definitely not going to work in NYC where people will just walk in front of a car if they know it will definitely stop, you need different infrastructure and legal frameworks.
AI-in-a-box servers, on-prem and in private clouds, make a lot of sense for privacy and latency. - FT
Emotionally intelligent robots are coming. Maybe 'robotic' will be redefined to mean 'smarmy AI drama queen'. - CMSWire.com
Prepare to get emotionally blackmailed by your psychologically manipulative chatbot - WIRED
Microsoft to talk about AI PCs, Copilot etc. at Build conference. - CNBC
Bill Gates plugs Sal Khan's book on how AI will improve education. - X (formerly Twitter)
In the latest OpenAI drama, Sam Altman says non-disparagement agreements were all a big mistake and OpenAI won't claw back equity. - X (formerly Twitter)
Brockman chimes in. - X (formerly Twitter)
Wolfram on AI in science and human progress. - Reason.com
Is ChatGPT getting smarter but less attentive to directions over time? - Harvard Data Science Review
LoRA parameter-efficient fine-tuning found to underperform full fine-tuning - arXiv.org
The guys who say, tech always increases productivity which increases wages are wrong on the theory. Historically true if labor and capital are highly complementary. But nothing in current models guarantees it. Labor share has been declining, and if capital becomes a near-perfect substitute for most labor then all bets are off.
Anyway Hinton is doing his job by giving fair warning. And the economists and policy wonks who are telling him, STFU and stay in your lane, it always works out, are derelict. It's a very disruptive thing and it will only work out if we monitor it closely and adjust, possibly dramatically.
Nobody knows. But it's not a given that tech makes everyone richer. Even if it does it's disruptive and there are big winners and big losers. And if robots are an increasingly perfect substitute for labor, we could get a singularity of robots making robots while the income share of labor goes down.
Life tip: To get Google to just give 10 blue links with no AI summaries, panels etc., go to Chrome settings, search for site search, add a shortcut called e.g. "Google 10 Blue Links" with shortcut 'gwo' to https://www.google.com/search?q=%s&udm=14 , then type 'gwo Sundar Pichai' in the search bar to default to web only.
Follow the latest AI headlines via SkynetAndChill.com on Bluesky