HN user
The feeling I'm getting is management took such a huge stance (the capex, the name change, etc) on it it'd be humiliating to turn 180 and ditch it now (they've spent 50 billion on Reality Labs so far). So they're quietly pushing it aside and one day when people aren't looking, they'll take it to the back of the house and shoot it in the head. Probably call it an "internal re-org" or something like that.
The only reason I can think that this atrocity got deployed was from an engineering perspective, creating AI profiles to engage theoretically gives FB a powerful lever to increase engagement, because you'd have this entire new field to collect data, experiment and test people's responses. In other words, it would turn engagement into a form tractable to optimisation algorithms.
I'd argue that it hasn't come to pass, but it is not wrong that its a better way to teach. Virtually all education has been set up in an industrial line style of pre-u education -> university -> companies -> profits -> some value to society. Every part has been attempted to be structured to optimise for the subsequent stage - companies optimisie for profits, universities optimise for employment rate and salary, high schools optimise for college entrance rates, etc. Yet because of Goodhart's and organizational incompetence, each part optimises badly, so education gets exponentially disconnected from each. His proposals are valuable to society, but the system is structurally against it.
The only way this works if it skips the assembly line right down to being of value to society, ie you set up a system which explicitly transforms that teaching into things like startups, research, nonprofits etc.
I don't think the Biden admin can criticize Xi for cyber warfare with a straight face.
In intelligence/performance. It's admittedly a fuzzy notion. Most benchmarks will probably show decreasing gains between generations. Similar to time/space complexity, trying to debate about what performance/intelligence is will get into a million definitions, caveats and technicalities. But a relative comparison between inputs and outputs is gives us useful information.
The inputs - data, compute and parameters - going into training these models have grown by many orders of magnitude between each gen. There's a lot of fuzziness about how much better each gen has gotten, but clearly 4 is not many orders of magnitude better than 3 by any reasonable definition. This mental model isn't useful to say how good each gen is, but it is quite useful to see the trend and make long term predictions.
We are likely going to get there. Similar to the steam/combustion engines (and other core technologies like computers, wireless transmission etc) there's first a massive rush to increase the power of it, at the cost of efficiency and effectiveness for more niche use cases. Then it is specialised to various use cases with large improvements in efficiency and effectiveness. My own prediction for where most gains will now come is
1) Creating new "harnesses" for models that connect to various systems, APIs, frameworks, etc. While this sounds "trivial", a lot of gains can come from this. Similar to how the voice version of ChatGPT was (apparently) amazing, all you really had to do was create an additional voice to text layer and another text to voice layer.
2) Increasing specialisation of models. I predict over time that end user AI companies (e.g those that just use models and not develop them), will use more and more specialised models. The current, almost monolithic, system where every service from text summary to homework help is plugged into the same model will slowly change.
From the amount of data each successive generation used (which grew many orders of magnitude each time) to the decreasing, logarithmic performance, it's quite clear the steam is running out on shoving more data into it. If one plots the data to performance graph, its horribly logarithmic. In another perspective, the ability of LLMs to transfer learning actually decreases exponentially the larger they and the data sets get. This fits into the how humans have to specialise in topics because the mental models of one field is very difficult to transfer to another.
I fear not the man who has written 10,000 features once, but the man who has written 1 feature 10,000 times
Something interesting I heard from a chip company that's huge in the space and work intimately with many automotive companies - why Chinese companies grow so fast is because their development cycle for a car is ~2-3 years, compared to traditional manufacturers who take 5-7 years. This is a massive edge in pushing out new features and exploiting the very rapid new tech - batteries, self driving, etc.
If one gets into the weeds of the whole messy thing, the Shanghai Communique acknowledges the position of both sides (Taiwan and China) that there is one China, the 1982 Joint Communique acknowledged the Chinese position and said the US had no intention of pursuing a "two Chinas" or "one China and Taiwan" policy. And of course there's a non stop stream of such political speak, which is a red herring for the whole issue. Beyond minor changes to wordings, the main things the US did that screwed over Taiwan:
1) Revoking the mutual defense treaty which legally bound the US to defend Taiwan, replacing it with a more vague law where military intervention was not clear. 2) Recognizing the PRC as the legitimate representative of China.
To align your claims slightly, Taiwan also claims the same section of the SCS (actually it claims a slightly larger part), so strictly speaking it is what happens if the "Chinese claims" are realised (Taiwan and China work jointly to support their claims, ironically enough). In real terms the economic effects of China with respect to the SCS are greatly exaggerated for a number of reasons. First being that shipping can be routed through Indonesia. It is not a chokepoint, it just happens to be the shortest route. Second, blockades have little to do with recognising swaths of ocean as territory. These are enforced by navies, and most blockades in history have not happened within the blockaders' own waters (for obvious reasons). It is no easier for China to blockade trade in the region if it claims the SCS. And third, of all the major economic powers, China has historically been the least likely to enact economic warfare like blockades or sanctions. There is also a fourth aspect where in the current political environment, the globalised, trade based economic order is the least popular in the US and assorted European states, not China (who in fact desperately needs trade).
There exists a third possibility where the US and China sign an under the table deal for China to invade, the US to saber rattle and China to allow the flow of chips to continue. The present direction seems to be the US is "de-risking" from Taiwan by moving chip production to the US, so if China does invade they aren't caught in a bind.
That is false equivalence, there's a class of things that should rightfully be discouraged from military use (chemical/biological weapons, land mines, phosphorous weapons) because of the significant harm/side effects they cause beyond some definition of acceptable.
That seems like a good idea. I am puzzled by what benefit the RL has in OP. It seems like a well defined constraint optimisation problem that could be done without RL, for example in the way you mentioned.
Google is getting enshittified. It's already visible in many small ways. I was just using Google maps and in the route they called X (bus) Interchange as X International. I can only assume this happened because they are using AI to summarise routes now. Why in the world are they doing that? They have exact location names available.
I'm also wondering this. Two possibilities. One, find the first principles/root node/glue that holds many disparate concepts together in some causative way. Two, the specific procedure/step/concept that you keep reusing across multiple problems.
The initial cram is an interesting concept. If you insert new things to learn at a constant rate, the repetition burden grows logarithmically. Assuming you have some fixed amount of time you can devote everyday optimally your repetition burden should be constant. So the solution is making new things to learn not constant, but front loading a lot of it.
I'm dubious of its scale and effects. Tech creates exponential value because of aggressive focus, iteration and critical mass of talent. The reserves seems very far from this.
I suspect the actual dollar value of AI from "companionship" far outstrips the value created from the endless junk like Gemini summaries.
Starting from the basics, Ramanujan was known to spend huge amounts of time in the library pouring over mathematical texts. He was also personally and spiritually obsessed with mathematics, thinking it was an expression of divinity. So its quite probable a significant chunk of his memories were already mathematical and random accesses to it were the same.
As much as I love Anki there is enormous friction in creating the cards, especially as a beginner. And you very easily get overfitting when you keep seeing the same question.
That is IMO already done by domain experts in the form of textbooks, courses, university classes etc. And with formal, tested relationships between various islands of knowledge. I'm not too sure how much I would trust an LLM to do this
Same for Anki, Notion, Zettelkatsen etc. Even ignoring the setup cost, it costs as much effort to insert in something as it does to retrieve it. The value prop tends to be low for individuals.
Perplexity has a dubious idea based around harvesting user chats -> making service better -> getting more user prompts. I am quite unconvinced that user prompts and stored chats will materially improve an LLM that is trained on a trillion high quality tokens.
The second idea being kicked around is synthetic data will create a new fountain of youth for data that will also fix its reasoning abilities.
I'd chime in on a somewhat lesser known part of networks that in practice the real determinant of performance: congestion control. The effects can be on the scale of an order of magnitude or two. The problem of how decentralised nodes in a network can use finite network resources optimally is fundamentally unsolvable (proved by Jaffe, intuitively nodes never have perfect information). Most congestion control algos use a congestion window (the amount of data in-flight) to limit themselves, this window is divided at a packet loss (aka congestion event) and increased when a successful round trip is made (in modern algorithms logarithmically if below a steady state assumption, cubic if above). There's a general consensus that these algorithms perform extremely poorly on long, fat networks because size of window growth depends on the round trip. Put another way, increasing bandwidth far from directly improves throughput. Latency in worst case is limited by the speed of light, and is in the order of the 150ms to 250ms range in the time it takes for light to travel alone. So there are fundamental limits on the latency that Starlink can provide.
Some back of the envelop calculations = it takes about 7ms in extra latency to travel up and down to the satellite vertically on a round trip, and modeling fiber optic vs starlink as two spheres of different radii starlink will take ~16% more time. Performance will be also extremely affected by the reliability of the medium ie packet loss from errors transmitting through the atmosphere vs fiber optic. Most congestion controls believe latency and packet loss comes from congestion and will aggressively back off. When I tested this on my own 866Mbps WiFi network with iperf and linux qc, with a 100ms latency and 1% packet loss my throughput was 1.86Mbps. At 5ms latency and 0.01% packet loss it was 58Mbps.
The number of cases AI is used for is too vast to cover, but I do have the same experience. I've a few heuristics about assessing how well AI will work.
1) AI's performance decreases exponentially with the sparsity of relevant examples available on the web. The fields in which these are prevalent tend to be specific frameworks, libraries or hardware. AI may sometimes extrapolate from similar examples - for example many JS libraries are similar to others - but when it comes to, say, firmware libraries, it fails horribly. How relevant is relevant is a matter of some dispute. I find that it cannot transfer much beyond direct examples of library code. It's not necessarily a matter of bad documentation, even if the necessary docs are there, it has very poor ability to come up with a solution from first principles.
2) AI's performance decreases linearly with the number of different dependencies that piece of code connects to. Part of it is it doesn't have the context, part of it is that the chance of it seeing that particular configuration decreases exponentially with the number of variable configurations.
3) Bizarrely AI is terrible at recommending books. I was asking it for books on introductions to category theory and it made a whole bunch.
Where AI is useful is writing boilerplate, especially starting code which has minimal dependencies and plenty of examples online.
Chinese startups have historically been the only ones to be able to move as fast as Musk. One very underrated ability of Tesla is they can develop new products in 2-3 years from conception to market. The only other car companies to do it at scale is BYD, Nio and Xpeng. The company he is talking about is probably LandSpace. They got from conception to hop test in what seems to be 5, maybe 4 years. SpaceX took 7.
Such a service is the opposite of a flywheel (a brake?) in practice. Those tokens are extremely low quality data.
1) I strongly double if even one in a thousand people actually bothers to report the outcome of their result. I and every programmer I know certainly don't (if we give up or it works we both just stop asking). And I'd think its one in a hundred thousand who gives detailed feedback that is useful for training. And one in ten million an expert that sits down and talks about its deep knowledge with it (and the problem of how you deduce this from chats that it isn't some crazy person).
2) AI's are extremely confident in their answer, which fools many people, especially those in a void of knowledge. Even if people did tell OpenAI whether every solution worked or not, I would heavily discount the accuracy of such data.
3) AI output autocannibalism does not lead to better outcomes, AI companies avoid using AI data for training like the plague. Mixed tokens I doubt would be much better.
The situation in reality is something like one in some huge number - maybe hundreds of thousands - of mixed tokens can be useful. Of those, those that are repeats of high quality sources like textbooks, dictionaries, man pages - have no value. Of the remaining, there is huge problem of how do you extract these needles in the haystack with high confidence. Given the incredibly lopsided confusion matrix (with the massive amount of actual negatives to actual positives) and this incredibly unstructured data set, I doubt its even remotely possible to find a way where you don't end up with a totally unacceptable ratio of actual to false positives. Letting this kind of garbage data in is how you get Gemini's gasoline spaghetti.
No truck on this robot vacuum race because I don't own one, but one an incredible name.
Counterpoint: How much of these novels are imaginative paranoia driven by cynical consumers? The closest similar incident - the Tongan castaways (granted they were aged 13 - 19 instead of 6 - 12) had the boys create a more or less exemplary society. They worked in teams, hunted and fish, created instruments to sing songs - no where even close to what happened in the novel.
I wouldn't exactly buy into the article. Its brilliant in that it puts a lot of things in your heads about Mark stepping back, but Meta's lobbying in dollars hasn't ceased one bit (its the 2nd largest lobbying spender after Amazon). Reading between the lines, Mark is preferring to use more traditional methods of influencing politics. Not that he is becoming neutral.