Xor and binary fuse filters require access to the full set of keys at construction time. In this sense, they are immutable. Alternatives have typically a fixed memory usage and a maximal capacity, but they also allow more flexibility such as progressive construction (adding keys one by one).
That's not an inherent property of the system. You can choose the most likely token(topk=1) and it will be deterministic (at least in theory, in some hardware setups it might be trickier)
I spin up a docker container using the docker API. I haven't used gvisor because I don't expect the model to try kernel level exploits. If it were the case, we're already doomed.
I can, and I would say it's a likely scenario, say 30%. If they don't have a significant edge over their competitors in the capabilities of their models, what's left? A money losing web app, and some API services that I'm sure aren't very profitable either. They can't compete with Google, Grok, Meta, MS, Amazon... They just can't.
I guess this students don't pass, do they? I don't think that's a particularly hard concern. It will take a bit more, but will learn the lesson (or drop out).
I'm more worried about those who will learn to solve the problems with the help of an LLM, but can't do anything without one. Those will go under the radar, unnoticed, and the problem is, how bad is it, actually? I would say that a lot, but then I realize I'm pretty useless driver without a GPS (once I get out of my hometown). That's the hard question, IMO.
As someone already said, parents used to be concerned that kids wouldn't be able to solve maths problems without a calculator, and it's the same problem, but there's a difference between solving problems _with_ LLMs, and having LLMs solve it _for you_.
Well the extent is much broader from a calculator vs an LLM. Why should I hire you if an agent can do it ? LLM is every job is a calculator and can be replaced. Spotify CEO stated on X that before asking for more headcount they have to justify not being able to do the job with an agent. So all the students who let the LLM do their assignment and learn basically nothing, what’s their value for a company to be hired ? The company will and is just using the agent as well …
An agent can't do it. It can help you like a calculator can help you, but it can't do it alone. So that means you've become the programmer. If you want to be the programmer, you always could have been. If that is what you want to be, why would you consider hiring anyone else to do it in the first place?
> Spotify CEO stated on X that before asking for more headcount they have to justify not being able to do the job with an agent.
It was Shopifiy, but that's just a roundabout way to say that there is a hiring freeze due to low sales (no doubt because of tariff nonsense seizing up the market). An agent, like a calculator, can only increase the productivity of a programmer. As always, you still need more programmers to perform more work than a single programmer can handle. So all they are saying is that "we can't afford to do more".
> The company will and is just using the agent as well …
In which case wouldn't they want to hire those who are experts in using agents? If they, like Shopify, have become too poor to hire people – well, you're screwed either way, aren't you? So that is moot.
So like arguably when people were not using calculators they made calculations by hand and there was a room full of people that did calculations. That’s gone now thanks to calculators. But it the analogy goes to an order of magnitude higher, now fewer people can « do » the job of many so less hiring maybe but not just on « do calculations by hand » but almost all fields where the use of software is required.
Where will all those new students find a job if :
- they did not learn much because LLM did work for them
- there is no new jobs required because we are more productive ?
Never in the history of humans have we been content with stagnation. The people who used to do manual calculations soon joined the ranks of people using calculators and we lapped up everything they could create.
This time around is no exception. We still have an infinite number of goals we can envision a desire for. If you could afford an infinite number of people you would still hire them. But Shopify especially is not in the greatest place right now. They've just come off the COVID wind-down and now tariffs are beating down their market further. They have to be very careful with their resources for the time being.
> - they did not learn much because LLM did work for them
If companies are using LLMs as suggested earlier, they will find jobs operating LLMs. They're well poised for it, being the utmost experts in using them.
> - there is no new jobs required because we are more productive ?
More productivity means more jobs are required. But we are entering an age where productivity is bound to be on the decline. A recession was likely inevitable anyway and the political sphere is making it all but a certainty. That is going to make finding a job hard. But for what scant few jobs remain, won't they be using LLMs?
> Spotify CEO stated on X that before asking for more headcount they have to justify not being able to do the job with an agent.
Spotify CEO is channeling The Two Bobs from Office Space: "What are you actually doing here?" Just in a nastier way, with a kind of prisoner's dilemma on top. If you can get by with an agent, fine, you won't bother him. If you can't, why can't you? Should we replace you with someone who can, or thinks they can?
You as the employer are liable, a human has real reasoning abilities and real fears about messing up, the likely hood of them doing something absurd like telling a customer that a product is 70% off and them not losing their job is effectively nil. What are you going to do with the LLM, fire it?
Data scientist and people deeply familiar with LLMs to the point that they could fine tune a model to your use case cost significantly more than a low skilled employee and depending on liability just running the LLM may be cheaper.
As an accounting firm ( one example from above ) far as I know in most jurisdictions the accountant doing the work is personally liable, who would be liable in the case of the LLM?
There is absolutely a market for LLM augmented workforces, I don't see any viable future even with SOTA models right now for flat out replacing a workforce with them.
I fully agree with you about liability. I was advocating for the other point of view.
Some people argue that it doesn’t matter if there is mistakes (it depends which actually) and with time it will cost nothing.
I argue that if we give up learning and let LLM do the assignments then what is the extent of my knowledge and value to be hired in the first place ?
We hired a developper and he did everything with chatGPT, all the code and documentation he wrote. First it was all bad because from the infinity of answers chatGPT is not pinpointing the best in every case. But does he have enough knowledge to understand what he did was bad ? And then we need people with experience that confronted themselves with hard problems and found their way out. How can we confront and critic an LLM answer otherwise ?
I feel student’s value is diluted to be at the mercy of companies providing the LLM and we might loose some critical knowledge / critical thinking in the process from the students.
I agree entirely on your take regarding education. I feel like there is a place where LLMs are useful but doesn't impact learning but it's definitely not in the "discovery" phase of learning.
However I really don't need to implement some weird algorithms myself every time (ideally I am using a well tested Library) but the point is that you learn to be able to but also to be able to modify or compose the algorithm in ways the LLM couldn't easily do.
>As someone already said, parents used to be concerned that kids wouldn't be able to solve maths problems without a calculator
Were they wrong? People who rely too much on a calculator don't develop strong math muscles that can be used in more advanced math. Identifying patterns in numbers and seeing when certain tricks can be used to solve a problem (verses when they just make a problem worse) is a skill that ends up being beyond their ability to develop.
Yes, they were wrong. Many young kids who are bad at mental calculations are later competent at higher mathematics and able to use it. I don't understand what patterns and tricks you're referring to, but if they are important for problems outside of mental calculations, then you can also learn about them by solving these problems directly.
Almost none of the cheaters appear to be solving problems with LLMs. All my faculty friends are getting large portions of their class clearly turning in "just copied directly from ChatGPT" responses.
It's an issue in grad school as well. You'll have an online discussion where someone submits 4 paragraphs of not-quite-eloquent prose with that AI "stink" on it. You can't be sure but it definitely makes your spidey sense tingle a bit.
Then they're on a video call and their vocabulary is wildly different, or they're very clearly a recent immigrant and struggle with basic sentence structure such that there is absolutely zero change their discussion forum persona is actually who they are.
This has happened at least once in every class, and invariably the best classes in terms of discussion and learning from other students are the ones where the people using AI to generate their answers are failed or drop the course.
> there's a difference between solving problems _with_ LLMs, and having LLMs solve it _for you_.
If there is a difference, then fundamentally LLMs cannot solve problems for you. They can only apply transformations using already known operators. No different than a calculator, except with exponentially more built-in functions.
But I'm not sure that there is a difference. A problem is only a problem if you recognize it, and once you recognize a problem then anything else that is involved along the way towards finding a solution is merely helping you solve it. If a "problem" is solved for you, it was never a problem. So, for each statement to have any practical meaning, they must be interpreted with equivalency.
There is a difference between thinking about the context of a problem and "critical thinking" about the problem or its possible solutions.
There is a measurable decrease in critical thinking skills when people consistently offload the thinking about a problem to an LLM. This is where the primary difference is between solving problems with an LLM vs having it solved for you with an LLM. And, that is cause for concern.
Two studies on impact of LLMs and generative AI on critical thinking:
How many people are "good drivers" outside their home town? I am not that old, but old enough to remember all adults taking wrong turns trying to find new destinations for the first time.
>How many people are "good drivers" outside their home town?
My wife is surprisingly good at remembering routes, she'll use the GPS the first time, but generally remembers the route after that. She still isn't good at knowing which direction is east vs west or north/south, but neither am I.
I'm like that too, but I don't think it transfers particularly well to LLMs. The problem is that you can just skip straight to the answer and ignore the explanation (if it even produces one).
It would be pretty neat if there was an LLM that guides you towards the right answer without giving it to you. Asking questions and possibly giving small hints along the way.
>It would be pretty neat if there was an LLM that guides you towards the right answer without giving it to you. Asking questions and possibly giving small hints along the way.
I think you can prompt them to do that, but that doesn't solve the issue of people not being willing to learn vs just jump to the answer, unless they made a school approved one that forced it to do that.
For your GPS at worst you follow directions road sign by road sign.
For a job without the core knowledge what’s the goal of hiring one person vs an unqualified one doing just prompts or worse, hiring no one and let agents do the prompting ?
Back in my day they worried about kids not being able to solve problems without a calculator, because you won't always have a calculator in your pocket.
Not being able to solve basic math problems in your mind (without a calculator) is still a problem. "Because you won't always have a calculator with you" just was the wrong argument.
You'll acquire advanced knowledge and skills much, much faster (and sometimes only) if you have the base knowledge and skills readily available in your mind. If you're learning about linear algebra but you have to type in every simple multiplication of numbers into a calculator...
> if you have the base knowledge and skills readily available in your mind.
I have the base knowledge and skill readily available to perform basic arithmetic, but I still can't do it in my mind in any practical way because I, for lack of a better description, run out of memory.
I expect most everyone eventually "runs out of memory" if the values are sufficiently large, but I hit the wall when the values are exceptionally small. And not for lack of trying – the "you won't always have a calculator" message was heard.
It wasn't skill and knowledge that was the concern, though. It was very much about execution. We were tested on execution.
> If you're learning about linear algebra but you have to type in every simple multiplication of numbers into a calculator...
I can't imagine anyone is still using a four function calculator. Certainly not in an application like learning linear algebra. Modern calculators are decidedly designed for linear algebra. They need to be given the rise of things like machine learning that are heavily dependent on such.
It's an interesting question! In my opinion, if you don't use tools it's very unlikely it can do any harm. I doubt the model files can be engineered to overflow llama.cpp or ollama, or cause any other damage, directly.
But if you use tools, for example for extending its knowledge through web searches, it could be used to exfiltrate information. It could do it by visiting some specially crafted url's to leak parts of your prompts (this includes the contents of documents added to them with RAG).
If given an interpreter, even if sandboxed, could try to do some kind of sabotage or "call home" with locally gathered information, obviously disguised as safe "regular" code.
It's unlikely that a current model that is runnable in "domestic" hardware could have those capabilities, but in the future these concerns will be more relevant.
Open source, fast and good: openrouter with opensource models (Qwen, Llama,etc...) It's not local but these is no vendor lockin, you can switch to another provider or invest in a gpu.
They could charge for tuning/support, just like every other Open Source company.
Most business will want their models trained in their own, internal data, instead of risking uploading their Intellectual Property into SaaS solutions. These Open Source models could fill that gap.
Personally i think it's cool but less interesting than cygwin and less interesting than wsl 1. A well integrated virtual machine is cool and all, but actually translating system calls or rewriting software for another environment is more interesting.
My understanding is they tried using more of a cygwin approach for a couple years with WSL1, but ultimately they weren't able to get some major software like Docker to work, not to mention the long tail.
WSL1 was much more interesting than cygwin: cygwin is a collection of ports. WSL1 used the existing binaries and transported kernel calls directly to a Linux kernel "personality" for the NT Kernel (which was designed to flexibly support multiple "personalities" like that way back in the day, but which hadn't been put to much use in recent decades before WSL1 brought it back and did something cool with it).
(WSL2 is a lightweight VM in a traditional VM sense.)
Interesting, sure, but WSL2 is simply much better to use as a daily driver and for production. I don't have to worry about whether stuff works or not like in WSL1, because, as you say, it's a VM so most stuff should work just fine.
I think everything except io is an okay in wsl1 if you only run user space software. But the io throughput is just bad. Iterate over 1000 image and hash each one costs only 1s in Linux and whole 20 seconds on wsl1. The way wsl2 implement fs is just far better.
Let alone if you use nodejs and have tens of thousands of files in node_modules. 'npm install' will cost you a whole 5 minutes.
WSL1 when accessing Windows files and Linux files uses direct IO via the NT Kernel. This is "slow" because NTFS has a different CAP theorem tradeoff than POSIX expected file system semantics. (It's direct file access so working with one big file is sometimes faster: the trick is that's what NTFS is better optimized for: bigger, fewer files atomic transactions. POSIX semantics work better for lots of small files and doesn't guarantee atomic transactions in the same way.)
From Windows (such as in Explorer) accessing WSL1 Linux files the safe way passes through a Plan9-derived file server as intermediary. This is surprisingly quick, but not without overhead. (But you can if you need to, do some unsafe operations directly on the files in NTFS.)
WSL2 when accessing Windows files accesses them through a Plan9-derived file server as intermediary. This is surprisingly quick, but not without overhead. WSL2 when accessing Linux files is using a Linux filesystem in a virtual hard disk file (VHD) similar to any other VM technology. Using a Linux file system it naturally exhibits POSIX semantics and is fast in the way Linux operations are expected to exhibit in lots of little files scenarios.
From Windows (such as in Explorer) accessing WSL2 Linux files passes through a Plan9-derived file server as intermediary. This is surprisingly quick, but not without overhead. Some operations Windows can do directly via VHD support in Windows.
The issue isn't NTFS as far as I understand (based on what the WSL team themselves have explained). The problem is that the NT kernel is simply slow at opening files. Windows has transactional NTFS but it's deprecated and hardly used. The slowness can't be fixed because the open codepath goes via a lot of different filter drivers that can hook the open sequence, combined with the fact that there's no equivalent of a UNIX dentry cache because file path parsing is done by the FS driver and not the upper layers of the kernel. Even if the default filters were fixed to be made as fast as possible - which is hard because they're things like virus scanners - third party filter drivers are common and would tank performance again.
It's a pity because Windows would benefit from faster file IO in general but it seems like they gave up on improving things as being too hard.
What I mean here by "transaction" semantics is not "transactional NTFS" (or the other distributed transaction engines that replaced it) but as a short hand for all the various different ways that file locking mechanics and file consistency guarantees are very different in NT/Windows/NTFS than in the POSIX "inode" model. All of that has a lot of complex moving parts (filter drivers are indeed one part of that complex dance both affecting and affected by Windows' file "transaction" semantics).
"Transaction" is a useful analogical word here for all of that complexity because how a mini-version of the CAP theorem trade-off space can be seen to apply to file systems. Windows heavily, heavily favors Consistency above all else in file operations. Consistency checks of course slow down file opening (and other operations too). POSIX heavily favors Availability above all else and will among other things partition logical files across multiple physical inodes to do that. Neither approach to "file transactions" is "better" or "perfect", they are different trade-offs. They both have their strengths and weaknesses. Using tools designed for one is always going to have some problems operating in the other. POSIX tools are always going look at Windows as "slow file IO" because it doesn't hustle for availability. Windows tools are always going to look at POSIX as willfully dangerous when it comes to file consistency. At the end of the day these filesystem stacks are always going to be different tools for different jobs.
Yup, nothing to magic, just the usual symptoms of Windows and Linux have always had different ideas of how files are supposed to work, so give Linux its own (virtual) hard drive instead.
I don't know anything directly about Microsoft's 9p plans, but the blogs give an impression they are considerably pleased at the 9p file server for what they've been using it for (especially these cross-platform communications) and they might use it for other things.
I really "like" or at best have mixed feeling of Linux/POSIX way of handling file in use, can be deleted/moved/edited, like EXCLUSIVE LOCK means nothing to the system.
Windows took a very different path from POSIX for a lot of reasons. It frustrates me sometimes when some Linux fans insist that POSIX file system semantics are "the best" and "the only right option" simply because they've been ingrained in more than a half-century of Unix development practices. The NT Kernel team was certainly aware of POSIX file systems and their trade-offs when they built the Windows file IO libraries and made different choices for good reasons. POSIX isn't the last word on how file systems should work, some open source developers should learn that, I think.
The reason they pivoted away from wrapping linux syscalls, etc. was that ultimately efficiently supporting features tied tightly to hardware (eg; CUDA) became extremely difficult (at least, achieving decent performance).
Virtualization is so efficient nowadays that it's much performant to go that route vs porting where often there will be difficult to debug performance regressions and bugs. So what WSL provides instead is much tighter integration between that linux VM and windows (including performant filesystem access, etc.).
It's a game changer because it's branded and marketed.
Soon people will be forced to use it in particular contexts, let's say for the DRM Subsystem For Windows Subsystem For Linux. And since you need the DRM Subsystem For Windows Subsystem For Linux to run those few pieces of crucial software, WSL becomes your daily driver. Then MS starts shipping their own downstream distro with even more extensions that hook into Windows...
DRM access I wouldn't be surprised by. But I highly doubt forcing their own distro. The whole point of WSL is being able to use off the shelf distros while staying inside windows.
Xor and binary fuse filters require access to the full set of keys at construction time. In this sense, they are immutable. Alternatives have typically a fixed memory usage and a maximal capacity, but they also allow more flexibility such as progressive construction (adding keys one by one).
reply