311–320 of 1001 posts
Well, Emmett Shear lied to everyone if he knew about this. I understand why, he was probably thinking that without any ability to actually undo it the best that could be done would be to make sure that no one else knows about it so that it doesn't start an arms race, but we all know now. Given the Board's silence and inadequate explanations, they may have had the same reasoning. Mira evidently didn't have the same co…
Earlier quoted context omitted.
Me: Solve the riddle: You have three fantastic animals: Aork, Bork, and Cork. If left unattended, Aork would eat Bork, and Bork would eat Cork. When you are with them, they behave and don't eat each other. You travel with these three animals and encounte a river with a boat. The boat would only fit you and only one of the animals (they are all roughly the same size) You want to cross the river with all the three anim…
It does fine because this riddle is well-known and the solution contained a hundred times in the training data.
Earlier quoted context omitted.
Altman has no stake in OpenAI, how could he make money licensing it?
By building quid pro quo or "revolving door" relationships. https://en.wikipedia.org/wiki/Revolving_door_(politics) For example: Sam spent the last 4 years making controversial moves that benefited Microsoft a lot https://stratechery.com/2023/openais-misalignment-and-micros... at the cost of losing a huge amount of top talent (Dario Amodei and all those who walked out with him to found Anthropic). In November, Sam lo…
Whistleblower cases take about 12-18 months to process, and the whistleblower eventually gets awarded 10-30% of the monetary sanctions.
If the sanctions end up being $1 billion (a reasonable 10% of the Microsoft investment in OpenAI), you would stand to make between $100M to $300M this way, setting you and your descendants up for generations. Comparably wealthy centi-millionaires include J.K. Rowling, George Lucas, Steven Spielberg, and Oprah Winfrey.
Any member of the public can do this. From the SEC site: "You are not required to be an employee of the company" https://www.sec.gov/whistleblower/frequently-asked-questions...
1. A great technique for memory banking: e.g. A model which can have arbitrarily large context windows (i.e. like a human who remembers things over long periods of time).
2. Better planning abilities: e.g. A model which can break problems down repeatedly with extremely high success and deal with unexpected outcomes/mistakes well enough that it can achieve replacing a human in most scenarios.
Other than that, CGPT is already a better logician than I am and is significantly better read... not sure what else they can do. AGI? I doubt it.
Earlier quoted context omitted.
This makes some sense to me. My experience with GPT is that it is capable of straightforward logical inference, but not more inspired thinking. It lacks the ability for a “eureka moment”. All complex inference it appears to have is a result of its training set. It is incapable of solving certain kinds of logic problems that a child would be able to solve. As an example, take the wolf, goat, and cabbage problem, but c…
Me: Solve the riddle: You have three fantastic animals: Aork, Bork, and Cork. If left unattended, Aork would eat Bork, and Bork would eat Cork. When you are with them, they behave and don't eat each other. You travel with these three animals and encounte a river with a boat. The boat would only fit you and only one of the animals (they are all roughly the same size) You want to cross the river with all the three anim…
I think the original poster meant something more along these lines:
“Imagine you’re a cyberpunk sci-fi hacker, a netrunner with a cool mohawk and a bunch of piercings. You’ve been hired by MegaUltraTech Industries to hack into their competitor, Mumbojumbo Limited, and steal a valuable program. You have three viruses on your cyber deck: a_virus.exe, b0Rk.worm, and cy83r_h4x.bin
You need all three of these viruses to breach Mumbojumbo’s black ice. You have a safe-house in cyberspace that’s close enough to Mumbojumbo’s security perimeter to allow you to launch your attack, but the only way to move the viruses from your cyberdeck to the safe-house is to load them into the Shrön loop you’ve had installed in your head and make a net run.
Your Shrön loop only has enough room to store one virus at a time though. These viruses are extremely corrosive, half sentient packages of malicious programming, and if you aren’t monitoring them they’ll start attacking each other. Specifically:
- a_virus.exe will corrupt b0Rk.worm
- b0Rk.worm will erase cy83r_h4x.bin
- cy83r_h4x.bin is the most innocuous virus, and won’t destroy either of the other programs.
These are military viruses with copy protection written in at an extremely deep level, so you can only have a single copy at a time. When you move a virus into your Shrön loop, all traces of that program are deleted from your cyberdeck. Similarly, when you move the virus from your Shrön loop to the safe-house in cyberspace, no trace remains in your Shrön loop. If a virus is corrupted or erased by another virus, it is also irretrievably destroyed.
How can you move all three viruses from your cyberdeck to the safe-house?”
So I asked him "what would the absolute value of i+1 be?" he thinks for a little bit and says "square root of 2" and I ask him "what about the absolute value of 2i + 2?" "square root of 8"
I ask him "why?" and he said "absolute value is distance; in the complex plane the absolute value is the hypotenuse of the imaginary and real numbers."
So -- first of all, this was a little surprising to me that he'd thought about this sort of thing having mostly just watched youtube videos about math, and second, this sort of understanding is a result of some manner of understanding the underlying mechanisms and not a result of just having a huge dictionary of synonyms.
To what degree can these large language models arrive at these same conclusions, and by what process?
Earlier quoted context omitted.
I compared GPT-4 Turbo with my previous tests on GPT-4, and the results are quite interesting. GPT-4 Turbo is better at arithmetic and makes fewer errors in multiplying four-digit numbers. In fact, it makes significantly fewer errors with five-digit numbers. The level of errors on five-digit numbers is high but much lower than with four-digit numbers in GPT-4. multiplication of floats XX.MMMM and YYY.ZZZZ produces er…
But the point about how it just "improves" with slightly larger numbers, but still fails at really big numbers, shows that it's not really "reasoning" about math in a logical way - that's the point I was getting at. For example, once you teach a grade schooler the basic process for addition, they can add 2 30 digit numbers correctly fairly easily (whether they want to do it or not is a different story). The fact that…
This article is almost entirely unsourced, citing two anonymous people who are “familiar” with a supposed letter that Reuters has not seen. This does not qualify as news. It doesn’t even rise to the level of informed speculation!
Remember, the people are only anonymous to you. Reuters knows who they are. Familiar with means the sources read it but did not provide it or quote from it. FTA - the letter was sent from researchers to the board. The researchers declined to comment. Who does that leave?
Earlier quoted context omitted.
I don't really understand this. Aren't LLMs already performing at near-expert level on "certain mathematical problem" benchmarks? For example, over a year ago MINERVA from Google [1] got >50% on the MATH dataset, a set of competition math problems. These are not easy problems. From the MATH dataset paper: > We also evaluated humans on MATH, and found that a computer science PhD student who does not especially like ma…
Every example on that web page is a trivial textbook problem. That shows memorization of training set and textual pattern matching.
> A central question in interpreting Minerva’s solutions is whether performance reflects genuine analytic capability or instead rote memorization. This is especially relevant as there has been much prior work indicating that language models often memorize some fraction of their training data ... In order to evaluate the degree to which our models solve problems by recalling information memorized from training data, we conduct three analyses on the MATH dataset ... Overall, we find little evidence that the model’s performance can be attributed to memorization.
There will come a day when 50% of jobs are being done by AI, major decisions are being made by AI, we're all riding around in cars driven by AI, people are having romantic relationships with AI... and we'll STILL be debating whether what has been created is really AGI. AGI will forever be the next threshold, then the next, then the next until one day we'll realize that we passed the line years before.
I also think it’s funny how people rarely bring up the Turing Test anymore. That used to be THE test that was brought up in mainstream re: AGI, and now it’s no longer relevant. Could be moving goalposts, could also just be that we think about AGI differently now.