So we have a situation when if AI hacks into something, nobody is responsible, but if you say something inappropriate to the AI, then you can be arrested. That is interesting ...
"In scary headlines like these, technology is the subject, and teens are passive victims." -- I am sorry, but how is this not true, when companies like Meta are in full control of their algorithms, tuning them to maximum addiction.
One can indeed argue what is the best way to deal with the problem of kids and social media, but I think the article is misrepresenting the situation.
I think old Android phones not supported by their makers are so problematic in the LLM era. I would think that even before LLMs the 3 letter agencies had exploits for those, but now one should assume common criminals will...
I am happy that the pixel phone I got has 7 year of support, but it is clear that Apple is in general is much better in this than all the Android providers (including google)
I think if there was a financial penalty of say 1% of profit/income for each "incident" where something was hacked/exploited, then the companies would be much more willing to think about safety.
While some of the text sounds sensible to me, i.e. some sort of external oversight + reporting of incidents, much of the rest seems focusing on limiting China and their open models as they are clearly approaching the abilities of Anthropic's models.
This really leaves a bitter taste....
"On Tuesday, September 1, we heard rumors that two Millennium Prize problems had been resolved. Inspired by these rumors and by the step change in performance of our internal model, we launched an effort to evaluate it on all open Millennium Prize problems and a few other high-impact problems."
IPO+rumour driven research.
I appreciate the achievement, but it doesn't feel right.
"Covariates included in this study were study center, child sex (female, male), birth order, mother’s ethnicity, parental highest education level and TV viewing time (hours/day), maternal age at delivery (in years), household monthly income at recruitment, and maternal sensitivity from a 15-min observation of mother–child interactions at 6 months. "
Unfortunately there are more and more cases where "The computer says no" and then there is no human to talk to, and you are bounced back from one automated system to another...
I had recently a case where my title (Mr) ended up as my middle name in the system of the operating airline (it was perfectly correct in the system of the airline that I booked the ticket through). Two hours of talking to those two airlines did not help me. I had to buy another ticket...
I think this starts to remind me the movie Brazil...
I don't think there is value per se, but it is just natural (at least in the environment I am in). I.e. when we discuss the analysis/code/ideas from one LLM or another, we describe it Claude/Gemini/Codex/etc did that and it comes like a person.
I don't find this take useful. We have little understanding of what conscience is, and how human train of thought really works. The modern LLMs IMO resemble more and more Chinese room problem. Maybe we don't like the mechanistic linear-algebra-based steps involved in the production of the output, but the end result is closer and closer to people's output. IMO this is very natural to start anthropomorphizing that.
If some text is AI written as a response to a much shorter prompt, then the prompt and/or sources used to make the text should be published instead (or at the very least together with the text).
The video in the post is very worth watching and is indeed scary. It is certainly true that it is in OpenAI's interest to publicize this, but I don't think the whole thing is invented. And seeing all this it is particularly scary if we think what will happen in organizations like NSA or similar in other countries. Presumably they happily adopt these techniques. And if you imagine a truly rogue state doing this, I can see an unimaginable damage happening very rapidly.
In the end, unless we are talking about art, in most areas I probably prefer something that follows some specific set of requirements and recipes, rather than have some sort of indescribable 'magic'. In that sense when it comes to coding AI gives you something that can follow the requirements pretty well and thus satisfies me in many circumstances. (but there are cases where human input is extremely important still)
Yes, but sometimes I don't need and want to learn. One example from my recent experience in research -- building custom dashboard pages for results of scientific analyses. Each analysis is bespoke, and building interactive webpages is simply not the skill many researchers have (and it's boring IMO). But here with LLM you could easily explore the results visually/share them with collaborators etc. There are plenty examples like that.
But certainly there are cases where learning is required.
I have just tried to switch to 3.6 instead of 3.5 in antigravity and it seems to constantly spit "critical instruction: STOP CALLING TOOLS NOW. YOU MUST WAIT FOR WAKEUP. ". I think I will switch back to 3.5
When it comes to coding, non-programmers do not have to be in a defensive position worried that their job is under risk, instead they just see a great tool that saves them time, especially doing boring coding like dashboards, visualizations, interactive web-pages, or doing experiments that they otherwise would not have time for.
In the end, standardization of this type of things can only be good even if the effect on waste is small. There is no need to create additional ways to confuse people.
The thing is that PG had introduced pluggable storage engines exactly for the reason I am talking about, and there have been few implementations of columnar storage using PG functionality, it's just they always stayed out of the tree.
So I wasn't talking about some functionality that is completely out of scope.
But I agree in the end I may be forced to move elsewhere...
PostgreSQL was a right tool for my task for many years. It is a question for PostgreSQL can adapt to a new reality of much bigger datasets or I have to switch to a new tool. And I am not the only user of Postgresql in this context. So it is easy to say in vacuum 'you are using the wrong database', but it's not something that can be easily changed with 100s of Tb of data, existing user workflows etc.
Speaking as long-term (>15 years) user of Postgres in science, I am getting worried about the lack of columnar type of storage in Postgresql. As the datasets become bigger and bigger, the limitations of PG's storage are becoming more and more significant. I know there are various extensions (i.e. cetus) that may offer such functionality, but then you depend on that extension being supported in the future, as well additional complexity.
Seeing the words "recursive self-improvement" I was expecting something else from the article. E.g. how the transformer architecture or agent design is being changed/improved through LLM automation, but the article mostly talks about the LOC counts.
When the consciousness itself not understood and well defined in the first place, it is pretty pointless to debate if something is or isn't conscious.
And here in particular the reasoning behind the argument is bizarre. Decomposing the complex activity into simple steps like 'predicting the next word' and claiming that surely can't have consciousness. A similar argument would be -- there is no way that movements of electrons by tiny distance would produce consciousness.
I guess you are happy with this: "Staffers there who may be political appointees and not necessarily subject matter experts sometimes ask for substantive changes in the research."