kromem

30 ms·

361.
by kromem 3y ago
Eh, it's going to end up moving to photonics. When we finally have NNs abusing virtual photons for the majority of network operations and using indirect measurement to train weights we'll have absolute black boxes performing above
362.
by kromem 3y ago
It only later developed into the general term as opposed to initially specifically referring to Epicureans. See Labendz, "Know What to Answer the Epicurean": A Diachronic Study of the ʾApiqoros in Rabbinic Literature (2003) And
363.
by kromem 3y ago
Would you bet that Oppenheimer would have by definition made better predictions how the bomb was going to change the future of war than someone who understood the summary of the bomb's technical effects from the scientists but also had
364.
by kromem 3y ago
Not really. So much regarding that topic depends on what's actually modeled in the training data, not how it is being trained on that data. They aren't experts on what's encoded in the training data, as the last three years h
365.
by kromem 3y ago
Right, but the question is if the engineers who intimately understood the function of the press itself were the experts that should have been looked to in predicting the sociopolitical impacts of the machine and the ways in which it would t
366.
by kromem 3y ago
I actually suspect the 'Thomas' addition was an archetypical figure modeled around the theme of twinning present in the aforementioned proto-Gnostic first century tradition, initially added to John (which was arguing against this
367.
by kromem 3y ago
I wouldn't be so sure of that outcome. There's diminishing returns on productivity gains at mega-corps that there isn't at small and medium sized businesses competing against them. If Walmart has AI that can automate 99% of l
368.
by kromem 3y ago
There's already a pretty huge one that people are sleeping on. A number of years ago I figured another way in which the might be evidence of being in a simulation outside of physics might be identifying any 4th wall breaking lore, as o
369.
by kromem 3y ago
In many ways, we're effectively discussing the accuracy by which the engineers of the Gutenburg printing press are able to predict the future of literature.
370.
by kromem 3y ago
"Feel good." The problem with nightshade is that while it worked in laboratory conditions, in practical application it wouldn't work without extensive coordination between everyone using it. For example, if one artist who dra
371.
by kromem 3y ago
Blake Lemoine. Google had a knee jerk reaction after he released the transcripts and got a bunch of press coverage. If you read the transcripts, it was a much more capable text model closer to OpenAI's products than what they eventuall
372.
by kromem 3y ago
I don't know - I agree with the xkcd that the period of AI development I find most scary is not when it is smarter than humans, but when it isn't yet smarter but improving all kinds of human tasks like warfare or aiding corporate
373.
by kromem 3y ago
The premise of annihilation presumes quite a lot. That the issues around continued training on only synthetic data without continued refresh for edge cases with organic data leading to model collapse will be solved, for example. Without tha
374.
by kromem 3y ago
In all likelihood, the opportunity cost of delaying AGI by a decade is at least tens of millions of human lives. A large part of why we think AGI poses an existential threat is because it's an anchoring bias around theorizing from the
375.
by kromem 3y ago
It's not "too intelligent." People need to stop thinking that the only way to get from A to B is the same way a human brain would traverse it. While you might be relying on significant semantic understanding and complex sen
376.
by kromem 3y ago
Cannibals don't necessarily kill the people they eat. For example, there's cannibalistic tribes that eat their recently departed ancestors as part of their tradition, but that's their only form of cannibalism. Cannibalism tec
377.
by kromem 3y ago
No, that's their intent. A lot of the LLMs have trouble with variations on typical logic puzzles. If you change an aspect, they will still regress to autocompleting the standard solution. In some cases, you can still get them to solve
378.
by kromem 3y ago
If I were any other AI company out there right now, I'd be circling Ilya like a vulture. Probably one of the sharpest minds on the subject right now and one of the few with an accurate picture of where the tech is headed and how to mak
379.
by kromem 3y ago
You suspect that my point is to weaken the voices of those who aren't anti-Semetic but question the actions of Israel? And not that perhaps my point is to highlight the opposite effect, by which the continued rhetoric of anti-Zionism a
380.
by kromem 3y ago
While that is definitely true and an important distinction, I will say that unfortunately all too often as discussions on the topic deepen there's a troubling correlation between the most vocal voices engaged in criticizing Israel and
381.
by kromem 3y ago
The overjustification effect doesn't occur if the extrinsic reward is a surprise or isn't known/expected in advance. So there are ways to improve it. I'm actually toying with going back into consulting for fun, in which
382.
by kromem 3y ago
Even if they aren't, there's a separate concern that we're past the inflection point of Goodhart's Law and this blind focus on a handful of tests evaluating a small scope of capabilities is going to be leading to model r
383.
by kromem 3y ago
The multimodal capabilities are, but the tone and insight comes across as very juvenile compared to the SotA models. I suspect this was a fine tuning choice and not an in context level choice, which would be unfortunate. If I was evaluating
384.
by kromem 3y ago
No, there's an odd disconnect between the impressiveness of the multimodal capabilities vs the juvenile tone and insights compared to something like GPT-4 that's very bizarre in application. It is a great example of what I've
385.
by kromem 3y ago
I would encourage people to be more aware of the overjustification effect. I used to lecture about it often when consulting, and despite that it still reared its ugly head in my own life. I used to love talking about tech. Because I loved t
386.
by kromem 3y ago
Right, because we all know the antivaxx member of the extended family never brings the topic up in person... People radicalized have their entire identity structured around a need for confirmation. When you only talk about a specific topic
387.
by kromem 3y ago
You didn't read more than the headline, did you? Directly from the abstract: "When individuals interact locally, the outcome is a stable plural patchwork of cross-cutting conflicts. By encouraging nonlocal interaction, digital med
388.
by kromem 3y ago
At no point have I ever thought "man, what corporations really need is a better chatbot." Chat as an I/O interface is very useful for AI companies, and the training data available by having many users interfacing with a LLM t
389.
by kromem 3y ago
It only leads to job loss if the labor demand is capped to the status quo. Which will actually be the case in far less situations long term than people seem to think. The reality is the early adopters who downsize staff to maintain current
390.
by kromem 3y ago
The smartest person I know turned down a position as CEO of one of the largest companies in the world to become CEO of a smaller and more independent oriented competitor and is apparently loving their job. In terms of developers, it's

More ›