Oh it very much depends, your opinion there is certainly not always true and not shared by everyone; you only know where you stand right now with any certainty.
It depends on the teacher. I’ve had good and bad math classes, and LLMs today are quite a bit better than the bad ones. The worst human teacher in my memory didn’t offer interactions. He walked in, turned his back to the class, wrote equations on the board for 45 minutes, and then left. The best math teachers, the ones better than LLMs, are the ones who share the joy and sense of discovery and history of math, and not just the mechanics. But there aren’t that many teachers of that sort.
Learning by using the internet without LLMs is rarely very good, but more often than not in my experience sucks much worse than using LLMs. If you include using public LLMs in internet usage, then it’s not very different from just using LLMs that search the internet. I have heard that a huge swath of today’s high school and college kids are reaching for chatGPT before Google (which is incidentally OpenAI’s goal), and that many of them would rather talk to chatGPT than talk to a teacher. I’m going to refrain from making any claims, but I believe there are a lot of people who disagree with your ranking of the options.
Foreign language learning is one case where I love using LLMs, because it’s not typically an option otherwise. You can practice non-stop and have conversations with someone fluent in a language who will be infinitely patient with your mistakes. This is true of math and other subjects too; using LLMs to practice, so that the human teacher isn’t the bottleneck, to supplement and reinforce the human interactions, is usually better than using the internet without LLMs. The other reason many people prefer talking to LLMs is the lack of judgement. If you aren’t getting it and ask the teacher one too many basic questions, they treat you differently. Sometimes it’s necessary and helpful, and sometimes it’s harmful and takes a long time to change. LLMs don’t do that, they just explain and explain. That lack of judgement is a big reason many people prefer LLM interaction to human interaction.
Okay, sure. The difference between option 1 and option 2 is completely irrelevant to this discussion and I gave what I thought would be an uncontroversial take on it because I thought it didn't matter. The point is at least one of option 1 or 2 is almost always better than option 3. Option 3 is what I and others are arguing against.
Irrelevant? How so? And why did you present two different options if you thought it was irrelevant? I thought you were trying to make a point about human interaction. From my POV, your options 2 and 3 are much closer together than 1 & 2, so if you’re not suggesting that human interaction is preferable to LLM interaction, then what’s your justification for claiming any option is better than any other? I still completely disagree with your clarified point, and suspect many many other people do too, and are demonstrating that preference by using chatGPT for their classes instead of using either the teacher or Google. It’s a simple fact that interacting with an LLM can be preferable to interacting with some teachers some of the time, and LLMs currently often make interacting with the internet more pleasant and more efficient. Your ‘almost always’ still sounds to me like one person’s opinion that is not shared by all.
We're not talking about humans vs LLMs in general, we're talking about humans copy-pasting LLM output as if they wrote it themselves. It's less "LLMs are bad" and more "if you want to communicate something specific to me, LLMs are a bad substitute for your own writing".
Oh I see I’ve misunderstood. Your analogy is restricting the scenario to no interaction with LLMs. In that case, there are still some reasons LLM writing might be prefereable for students: for non-native speakers, and for math teachers new to match developing a new curriculum (sucks, I guess, but it certainly happens), just to name a couple.
One question to ask is why is your contrived LLM scenario any different than a math teacher using a textbook, or district/state worksheets? Math teachers typically do not write the course text or exercises, and never have. Very few math teachers do their own writing.
I’d guess the concern is performance, not what initializer value is used. And performance is a valid concern that is discussed in the proposal, and a reason there’s an escape hatch. Still, it might cause some confusion.
They were saying the problematic philosophy started in C++ 20, not the variable initialization rule.
Yes the reason is obvious, but it’s neither simple nor black and white. One huge problem is that this can cause serious performance regressions, and you have to change your code to opt out, e.g. add “[[indeterminate]]”. There are many, many cases in high performance computing where the intended & desired behavior is don’t touch my variables until I fill them.
This is changing C++ core principles, there’s a new designation for the state of a variable: erroneous. It’s also subtle and weird, because you can still have well-defined behavior even with erroneous state. It does seem like this might be an experiment though, I don’t think this is the end of the story. (It seems they’re already talking some redesign of this idea.)
What I'm most annoyed at with the variable initialization change is that:
- It's potentially a performance change in every single function, especially ones that have sizable fixed-size buffers
- If you have regressions you have to spray [[indeterminate]] everywhere, because there is no coarser way of suppressing it.
- While the language says unrecognized attributes are ignored, compilers frequently warn on unrecognized attributes. Clang, for instance, currently warns on [[indeterminate]].
- There is no defined macro name for backwards compatibility.
Which means that libraries are going have to all declare their own macros for [[indeterminate]] and pepper their code with it.
Uninitialized variables were already UB to read, because some architectures have trap representations, even for integers. Every register on Itanium has one.
That's assuming you were reading it without writing to it. There are three common cases when that isn't true.
The first is that you have a fixed buffer large enough for the maximum message size even though the typical ones aren't that big. You most often write 1% of the buffer and read it back, the other 99% is never accessed.
The second is that you always write the entire contents before reading it but the compiler may not be able to see that.
And the third is that you have a code path where that variable is simply not used.
You would then have the compiler emitting instructions to write zeros that are either overwritten before being read or are never read at all.
Moreover, zero initializing the data doesn't actually remove the bugs when that isn't the case. Consider the first case when you mess up. You have a fixed buffer used to store variable length messages. For the first message the buffer is now zeros instead of uninitialized, but for every subsequent message the remainder of the buffer still contains the remainder of the previous message and subjects you to information disclosure or data modification if you're reading back a different amount than was written in the associated call.
Now consider the second or third case. You unintentionally read from a variable before assigning to it. You get zeros instead of uninitialized memory, but if you weren't expecting zeros, well, the UID field is now 0.
If you are writing to it before reading from it then it's not uninitialised, and in 99% of cases the compiler can see that and will not set it to 0 at its declaration because that's a dead store.
The function maybe_fill_buffer() is an external library function that either fills the buffer and returns true or doesn't access it and returns false. Or maybe it unconditionally fills it, or unconditionally returns false without reading from it. The compiler can't see any of that though because it's in an external library. For all it knows that function is going to read from it instead of writing to it.
Notice that if it could actually figure it out 99% of the time then it could also emit a warning the 1% of the time that it can't and encourage you to make an explicit choice, which would have been a better option if that was actually the rate.
Three of those four prompts are opt-in and configurable and aren’t the site’s fault, no? I disable the browser and OS password managers because I use 1Password… why do you have the Chrome and OS password managers enabled if you’re using 1Password? Or if there’s some reason you need them all available, at least rather than dismissing the prompts, why not add an entry to disabling prompting for a specific site in the 2 managers you don’t want to see again? All three support that.
On my phone I don’t see any way to disable the system passkey option. I have only 1Password toggled on in my settings but I still see the system prompt to save a passkey. It’s maddening. I’ll have to double check on desktop but I thought I had toggled off everything but 1Password.
Perhaps I missed something or perhaps it really is that user hostile. It wouldn’t surprise me in the slightest.
If any of them are configurable, I'm not smart enough to figure out how. I looked again, and I'm still not seeing anything. Even if it's my fault I'm seeing these, I still resent them.
There’s not much evidence these are being used, only that they are dependencies for something else; that’s why the download numbers are so high.
I wouldn’t say it’s broken, I’d say there are tradeoffs, and devs have known this and discussed it since the start of npm or any package manager. You automatically get some bloat when you use other people’s software. That’s the downside. The upside is you don’t have to write the code yourself and you can create things more quickly by not solving problems that others have already solved.
It’s worth noting that AI has some of the same tradeoffs. The quality of what you get is still proportional to your prompting & reviewing effort, and spending low amounts of effort often results in similar amount of bloat.
> Cracks are starting to appear. Open Ai and Anthropic are publicly asking for slowdown in AI research. Translation: We see this technology not being any more useful than what it is now, no AGI is coming
Predicting a usefulness plateau is absolutely wild given how fast AI agents have been improving at writing code this year. I have doubts about AGI but I think you’re making assumptions and translating it wrong. These two companies have always been asking for a slowdown from their inception, that’s not a new thing. It’s part marketing hype, but they both do want regulation to step in and slow down the competition, not because they see a usefulness plateau, but the opposite - the usefulness is growing so fast that they want to remain in control, and they are scared that working hard and competing will not be enough. Anthropic has also said out loud they think their competition (not just OpenAI) is not being responsible and they want the regulation so they can be the responsible shepherd, as AI gets more and more useful.
For me opus-4.8 was the most useful coding model. Sure fable is better at planning but is expensive to use as a daily driver. But even fable, when it fails, fails in such strange ways that I am now convinced this intelligence is an illusion and path to AGI lies elsewhere. I was willing to buy the whole emergent intelligence claim till last year. Now, not so much.
As for calling for slow down, sure the two companies kept parroting each other's lines but they were not slowing down the cash burn or gpu purchases. Why would they? The prize was too high. Now, with open Ai needing trillion dollar valuation to IPO and private funding possibly showing signs of slowing down (only for these labs because the valuation is too high to begin with for most prudent investors: My speculation) they have no option but to slow down. Then would you rather say, slowed down because we are running out of cash or that we are slowing down because national security? It really cannot be that they can't solve alignment but otherwise it is really powerful and improving. Simply because airgap exists and we know how to do it. If all else fails power off the freaking gigawatt cluster. More likely this recursive self improvement is an unstable loop and the model is likely degrading with self improvement effort. At 100's of millions per experiment this is going to be unsustainable.
Ah, now having the best model be expensive to use, and thinking the cash burn of both companies is insane and unsustainable I completely agree with. This, I think is likely the biggest reason they’re asking for government intervention and regulation: to help them weather the coming investment/cash plateau, not because the models are nearing any asymptotic limits. To be fair, there is an argument to be made that the improvement in models is tied to the cash burn; if they can’t train bigger models or acquire more data or research more effective harnesses, then the improvement of their models might slow down - while GLM or other models continue to improve.
I’ve never thought LLMs were on the AGI path, but I have to admit it’s surprising how far it’s come with no end in sight yet. There is something important to be said about how ‘intelligence’ is embedded in language, and it suggests that intelligence isn’t exactly what we thought it was. The language component of intelligence also goes a long way to explaining technology’s progress in human civilization; how language and the printing press and mail and radio/tv and the internet have each ushered in accelerations in the pace of progress. Biologically and evolutionarily speaking, it’s unlikely that humans have become any smarter in the last two thousand years, but technology (among other things) has exploded.
Astra, on average, despite its GPT 6 version bump, is not any better at coding than Sol. Some even argue its worse in practice due to the varying quality of its output.
(^ this "swearing at a model" thing has happened to me multiple times on Astra already)
If you have to "debate" the quality of a new Big Number model (and double and triple check your eyes and model setting switches when it pukes up complete garbage), that is NOT a good sign.
There is a theory that they are training them on those benchmark tests. No way to know if that's true, but there is also no reason to trust these companies to not do something like that. They lie a lot.
The "how much time does it save a dev" tests seem a stronger way to measure success, but not seen one of those run for a while.
Not bothered with Astra myself, but the demos I've seen people build don't seem any more impressive on the important stuff. Defaulting to three.js for games just feels smoke-and-mirrors to make them look better, as those games are still as unplayable for the same reasons they were in 2D.
I agree with you generally, just an observation on coding specifically.
Have the models improved since Opus 4.x? I find the newer models are not better in my day job, maybe in one shotting mvp's and other tasks.
Not trying to argue your point, just intrested in the coding aspect, if the models were improving as fast as benchmarks I would expect capability improvements to be obvious, but talking to people and reading forums, it seems everyone has a different opinion.
Yes, benchmarks are gamed and only loosely indicative of real world performance.
Also yes, Fable is massively better than Opus. It requires significantly less instruction and specs and produces more directly mergeable code.
The improvement is obvious as soon as my Fable allotment runs out and I try to do something with Opus. Have you given the same (larger) task to Opus and Fable?
Sure of course there will be some internal reason or some non-zero value behind doing any craft. But the analogy to AI & software engineering careers is that there are corporate for-profit tomato farms today. They have pushed out plenty of small farm businesses. Your mom enjoys tomato growing in spite of the availability of cheap high quality tomatoes. Your comment seems to me to agree with and support the GP comment entirely; you mom found value in doing the work despite there being no financial incentive.
Many people have argued over the years that it would be great when we have an economy that takes care of all the labor and everyone in society is free to choose to spend all of their time on things that have no financial incentive.
That would be great, if everyone in society thought about each other as altruistically as those pushing for UBI and stronger safety nets. It's probably achievable in small homogenous countries, not in this vast, heterogeneous, central-government-wary USA.
Aside from that not being what was said (the parent’s pull-quote is wrong in multiple ways), it is in fact not possible to tune a piano to any arbitrary set of 88 pitches without changing the design so dramatically and fundamentally that it would no longer be a piano. The triangular shape of a grand determined by the strings lengths, the placement of the pins, the use of double/triple strings in some frequency ranges, and design of the sound board, all permit only a very limited range of frequencies for each key on the keyboard. You very obviously cannot take an existing piano and tune it to 88 keys of A0, nor to 88 keys of C9.
This is a great way of summarizing a lot of the TV we’re getting in the US that have British characters, and it’s annoying once you notice it.
See how the accent changes in Lie to Me between season 1 and season 3:
S1: https://youtu.be/bWyhsqh_e9s
S3: https://youtu.be/oPqOET_xCKw
reply