I'm surprised by the pushback by some people here; I wonder if they're actually learning French (or another language) themselves or just reacting generally against use of LLMs vs. humans here (a common trend on HN). I am learning French, and LLMs have proven to be very useful for me. Great to see the author go beyond current LLM voice exchanges by employing spaced repetition; thank you! I'll check it out.
I learned French, now living in France. I took a course to pretty much improve my accent and to not rely on translation when speaking. Our teacher never spoke English to us since day one, using images instead to relay messages.
What really helped me enforce what I learned is talking to real people online. There are plenty of groups in Discord where you can have conversations with people of different backgrounds and personalities, where you can capture vocabulary used by French speakers every day.
LLM is like Duolingo, you cannot rely on it alone. The OP cancelled their French course for this. You can use it but make sure to have different active and passive learning methods.
There is a point where you just need to start talking to actual people. Spending 6 months in a country that speak the targetted language works wonder for example once you know the basics.
Aren't there language exchange meeting groups in your area?
Most of the German discord group is mostly annoyed at how awful LLM explanations and hallucinations are.
Communication and teaching is something that teachers and tutors are far better at than LLMs. In fact, most seem to agree that the
30EUR or $40 textbooks with well organized listening/speaking tests are leagues better than any LLM subscription. And it probably will take you months+ of daily work to go up a language level. (a2 to B1 or B1 to B2), if not a year++ if you are more casual at learning...
At best, LLMs are a tool for browsing the free web for other resources.
As a teacher they have several flaws:
1. They understand your broken grammar and work with you -- bad. Real native speakers will struggle with bad grammar and pronunciation. You the human need to feel this constantly so that you know where to improve. Feeling the instinctive disgust from the other human is part of what helps us know what to practice.
2. They fail at coursework. A textbook puts you on the proper course, already graded to the level you are on and with exhaustive layouts of the subjects you are expected to know.
3. They understand your broken grammar and converse with you without a full ability to explain why it's wrong. Yes, this really is big enough to mention twice.
----------
I'd say LLMs are a reasonable tool for maybe finding additional grammar resources (ex: another source explaining N-declension, or other subjects you know you are struggling with). But as a general guide??
They don't know what you don't know. You still fall into the beginner trap of spinning in circles. If you already know what to search for, LLMs can accelerate the process but in my experience all the already available cheap textbooks are better sources of exercises and graded listening/reading material.
-------
IE: if you want to get A2 graded German reading and listening, get something like Hueber Lesehefte / Sicherheit ist nur wins Carsten Tsara blickt nicht durch for like 10 EUR (.mp3 read along and .PDF).
Or the myriad of other graded readers already available.
Your test is simple. Just read and understand the book. It's A2 after all, if you believe yourself to be at A2 level (or are aiming to achieve A2), then just read A2 stuff constantly.
It's just the usual LLM stuff. Beginners get wow'd by the chat bot interface but ignore all the issues with learning with LLMs. And beginners also find the well trod path of textbooks and study to be boring. But at the end of the day, the textbook is a simple and consistently useful tool, while LLMs arent.
Oh, I'm not talking about using only LLMs, for sure. I am currently taking a college sequence (currently French 3), with textbooks, homework, and everything. 100% agreed on the value of traditional texts/flows. On the matter of LLMs being permissive with grammar, you need to use the tools specifically meant to check grammar. I use LanguageTool for everything I write to make sure my grammar is both correct and current.
LLMs are invaluable IMO for immediate feedback on your writing and speaking. I've started using French inputs for most regular interactions, even with things like Claude Code for my work and for all LLM interactions (I use Grok). Before I input into Claude Code, I check the grammar and content with LLMs. Tutors and traditional paths will not help with this.
The other sequence I'm starting to do is to have conversations as I take baby steps in speaking/listening. Agreed that talking to real French speakers is the best, but initially we tend to be shy because we're so bad, and using a non-human can put us at ease. You may be different, but IMO this is typical. And this post was about refining this via an app that adds spaced repetition, and I appreciate that.
Have you tried looking for a French Beginner Discord voice chat? Or a similar local club for an in person meeting in your area?
Tutors are not necessary. But this is a humanity / language arts problem. Communication with humans is literally the goal. Conversing with other similarly ranked beginners (with a moderator or teacher guiding the group) is among the best practice you can do.
Again, the point of public discussion with other humans is to find all the little mistakes and misunderstandings. The things that other humans find difficult with your pronunciation, rhythm or accent.
Tutors are simply the most expensive version of this available of live, one on one practice. It's a bad overall $$ value compared to class discussion or clubs but the human in the loop is perhaps the most important thing here for training yourself.
Again: it's the LLMs overly generous acceptance that I find problematic. When I talk in German at A2 level, other humans easily point out when I'm doing things in "English order" or other similar mistakes (which makes the sentence harder to understand in German).
But with an LLM, the LLM just understands broken German in English order or whenever I leave out a separatable verb or whatever. It autocorrects too much.
I'm no longer sure you have to, actually. I mean, we do trust the assembly that compilers produce without having to read it, don't we? We're rapidly getting to that stage with LLMs, IMO.
The assembly is a deterministic transform of the input logic, and if it doesn't match then it's a bug in the compiler. If an LLM-based code generator doesn't match what you asked for, that's OK, just pull the slot-machine handle again. that's the difference.
The "pull the slot-machine handle again" is the dangerous thing here.
I can feel it sometimes, as my brain shuts down and I gamble instead of thinking. It's a reversion to what I call "monkey mind" where you just keep pressing buttons to "make it work". I took a decade training my mind away from this, and too much AI is bringing it back.
The issue here is not just the gambling aspect, but verification that you got what you asked for. Code problems can be extremely subtle, and the cost of verification is deceptively high, ruining the advantages of using AI in the first place.
I truly believe that people claiming huge productivity gains from AI are either terribly slow programmers or are skipping their due diligence. Many "vibe coders" are incapable of checking the output of the code.
And then getting bugs when they use a new version of the AI, just like people occasionally got bugs when they upgraded to new versions of the compiler...
they would get bugs on every invocation of the software, not on a new version of the AI. it's equivalent to your compiler have a RAND function in it where it chooses between a billion different options every time it compiles, it's absolutely not equivalent to a compiler having a bug.
They’re not, and will never be in their current form and architecture.
Compilers are mechanical and engineered to produce a correct output. A compiler emitting incorrect machine code is exceedingly rare, and considered a bug. They have heuristics and probabilities in them, but those are to pick between a set of known-good outputs.
An AI is a bag of weights outputting a probability of the most plausible token that follows [1]. It is inherently probabilistic in nature and its output is organic (by design, they’re designed to mimic human speech), as opposed to mechanical like a compiler.
A compiler follows hard rules. An AI does its best.
And to be fair, AIs are no better than human in this regard: humans are pretty bad at generating correct code without mechanical tools to keep them in line (compilers, linters, formatters). It’s not a wonder we use the same tools to keep LLM output in line as we do humans. (And, to be fair, LLMs are better than humans at oneshotting valid code).
[1]: to those that tell me this vision of an LLM is outdated: nope. The heavy lifting is done in the probability generation. Debates about understanding are not relevant here, and the net output of an LLM is a probability vector over raw tokens. This basic description can be contrasted to a compiler whose output is a glorified Jinja template.
Not OP, but it means nothing, because it's not "effectively" becoming a compiler.
Think about it from an information theory standpoint:
A compiler takes at least the exact amount of information it needs to produce a result, and produces exactly that result every time (unless it's bad at its job or has a bug).
An LLM always takes far less information than would actually be needed to fully describe the desired output, and extrapolates from that. It fetches contexts and such to give itself a glut of assumedly relevant information, but the prompt always contains less information than necessary to produce the code it generates. If it did fully contain enough information, then you've just written a far more verbose version of the program in human language.
Yes, I am saying there's no functional difference (for practical purposes) between a deterministic transformation like a compiler and a perfect probabilistic transformation like an LLM.
We do not have "perfect" probabilistic transformation, and we probably never will (in part because it's hard to know what exactly that even means), but the gap between the two is shrinking every day.
Ergo:
> they're becoming (effectively) more and more similar every day.
I know it’s tiring to talk about “hallucination”, but truly, models still do hallucinate
They constantly say they did a thing they didn’t, say they know how to solve something when they don’t, etc. Regardless of guard rails or tests - AI forces a constant vigilance of a new kind.
Not just “what might have gone wrong” but also “what do I think is working but isn’t actually”.
And we’re not even talking about how it chooses substandard solutions, is happy to muddy code/architectures, add spaghetti on top of spaghetti etc.
Agentic coding often feels like an army of unexperienced developers who are also incredibly eager to please.
"Still" means "it always had hallucinations, and it still does, despite people thinking that it doesn't anymore". People think we've moved past that. We haven't.
This is a really, really, really bad comparison. I used to say the same thing. But the semantic distance between compiling a for loop to equivalent assembly instructions is much smaller than the distance between "I'd like a web application that can store and retrieve todo items." The space of the latter is practically infinite in what can be "compiled."
A counterpoint, since I never made that logical jump in your latter part of your comment: programming languages are, functionally, all domain-specific languages and do a good job of either describing directly, or consistently, deterministically, providing a reasonable and unambiguous abstraction over low-level concepts expressed by assembly languages.
Human languages are mostly very bad at this, and in particular bad at mapping low-level abstraction to the human written word unambiguously in a way that is as expressive as programming languages.
Inference closes that specific gap significantly (which is why anyone at all sees LLMs as a useful option to explore), but it will never be as good as a purpose-built language designed to map to a reasonable corresponding assembly language implementation.
> we do trust the assembly that compilers produce without having to read it
Yes, because wrong assembly blows really loudly. From wrong behavior to invalid instruction errors and everything between them. Moreover, compilers are battle tested over the years, with extremely detailed test suites, and extreme testing (everyday, hundreds of thousands users test and verify them).
Also, as people said, assembly generation is deterministic. For a given source file and set of flags, you get the same thing out. Byte by byte, bit by bit. This is what we call "reproducible builds".
AI is not like that. It's randomized on purpose, it pulls from training set which contains imperfect, non-ideal code. "Yeah, it works whatever", doesn't cut it when you pull a whole function out of its connections, formed by the training data. It can and will make errors, because it's randomized from a non-ideal pool.
Next, sometimes you need tight code. Fitting into caches, running at absolute performance limit of the processor or system you have. AI is not a good fit here. Sometimes you go so far that you optimize for the architecture at hand, and it works slower on newer systems, so you need to re-optimize that thing.
For anyone who reads and murmurs "but AI can optimize", yes, by calling specific optimization routines written by real talented people for some cases; by removing their name, licenses, and context around them. This is called plagiarism in its mildest form and will get you in hot water in academia, for example. Writing closed source software doesn't make you immune from cheating and doing unethical things.
Lastly, this still rings in my ears, and I understood it over and over as I worked with more high performance, correctness critical code:
I was taking an exam, there's this tracing question. I raise my head and ask my professor: "Why do I need to trace this? Compiler is made to do this for me". The answer was simple yet deep: "If you can't trace that code, the compiler can't trace it either".
As I said, I just said "huh" at the time, but the saying came back and when I understood it fully, it was like being shocked by a Tesla coil.
Get your sleep, eat your veggies and understand your code. That's the four essential things you need to do.
I've actually taken to double-checking the assembly in some instances. There are surprising times that the compiler won't make the shortcuts and optimizations you thought it should, and I also used this method to call out an unsuitable compiler since I caught it spitting out some ridiculous 10x-long set of instructions in certain critical instances.
We are not rapidly getting to that stage with LLMs and frankly it's hilarious that you are claiming so.
For anything other than Greenfield, new code projcets without dependencies and conventions and connections to other proprietary code, it has to be reviewed. Even for that case it's not good to not review code
>AI safety and wide distribution of benefits are the core of our mission. Two of our most important safety principles are prohibitions on domestic mass surveillance and human responsibility for the use of force, including for autonomous weapon systems. The DoW agrees with these principles, reflects them in law and policy, and we put them into our agreement.
So they're saying anthropic is lying or what? Because Sam Altman is saying that DOW agrees with no mass surveillance and no autonomous drone killing.
Also if not, how safety is their priority?
Sam Altman is lying by omission. It’s been confirmed [1] that OpenAI agreed to "lawful" use of the models. Since it’s the DoW, they can make pretty much anything lawful by invoking "national security".
We pay 7.25% sales tax in California, the highest in the country. Plus, county taxes can go up to 3.5%, adding up to 10.75% total. It's not too much to ask for basic stuff like maintaining the streets when paying this kind of money. The roads in Orange County, where I live, are great though.
That's not too far off from the total sales tax in -say- one of the largest metros in Alabama; Birmingham. Total sales tax in that city is 8.0%. [0] I can tell you from personal experience, that you get a lot, lot less for that money than you do in California.
Whoops. If we're just comparing state sales tax, CA does seem to be the highest. However it's not dramatically higher than many other states, and when you throw in local taxes the combined rate seems to be just in the top ten or so.
The problem is that biases tend to be built in via even rudimentary stuff like bad training material and biased tuning via system prompts. E.g., consider the 2026 X post experiment, where a user ran identical divorce scenarios through ChatGPT but swapped genders. When a man described his wife's infidelity and abuse, the AI advised restraint to avoid appearing "controlling/abusive." For a woman in the same situation, it encouraged immediately taking the kids and car for "protection."
The bot was trained on conservative bullshit. In this scenario, woman taking the advice would end up punished by court. And that happens even when there is documented history of domestic violence in play.
Articles like this keep popping up because they're catnip to those who hate AI or feel threatened by it. On coding, I routinely see people trashing vibe coding and jumping on the slightest mistake agents may make, never mind that human devs screw up all the time. And write-ups citing stats on AI coding tend to be written by folks who either don't code for a living or never earnestly tried it.
I use Claude Code regularly at work and can tell it is absolutely fantastic and getting better. You obviously need to guide it well (use plan mode first) and point to hand coded stuff to follow, and it will save you enormous amount of time and effort. Please don't put off trying AI coding out after reading misinformed articles like this.
I think devs have a natural inclination to resist a seismic shift in their industry, which is understandable.
However I agree that a lot of this stuff is FUD and AI dev is like a new skill, it takes time to master. It took me a few months but I’m comfortably more productive and having a more fun time at work with Claude Code
There's a BIG difference, at least with tools like Claude Code: plan mode. I'm now using Claude Code a lot at at work, and the first thing I do is enter plan mode where I can have a "conversation" asking it explain how it would implement. Just a few back/forth later I end up refining its plan to conform to good (or what I think is "good") design, after which it will tell me exactly what it is going to do (with code diffs), which I sign off on (again, potentially after a few iterations). It's only then that it generates the code.
By contrast, on one project many years ago I was reviewing the code generated by an overseas team, and I couldn't make head or tail of it, it was an absolute tangled mess that was impossible to fix.
Well, I'm sure we've all seen code produced by human developers that is 10x worse than what my Claude Code produces (certainly I have), so let's be real. And it's improving scary fast.
This seems like a lack of experience. The more I work with LLMs, the better I get at predicting what they’ll get wrong. I then shape my prompts to avoid the mistakes.
I think the bar has raised, for sure. There's code I work on from prior seniors that is worse than what our current juniors write, I'm assuming AI is assisting with that but as long as the PR looks good, it's no different to me.
I've noticed that generally OK design patterns and sticking to idiomatic code has increased while attention to small but critical details remains the same or maybe slightly decreased.
Hard disagree. Humans fail in ways I know, can predict, and know where to look for. ML coding assistants fail in all sorts of idiotic ways and thus every damn line needs to be scrutinized.
What actually scares me is the idea that with humans you can manage to follow their train of thought. But if LLM just rewrites everything each time, well that is impossible to follow and then there is same work to be done over and over again each review.
Questioning science is not automatically "anti-science," IMO it's best to remain skeptical and stay focused on the evidence. The fact of the matter is that current "best medical advice" is not the best either in terms of quality of life or prognosis. I've had a remote member of a family lose sight in an eye, develop short term memory issues, and rapidly deteriorate from cancer in spite of following the best medical advice and guided by top physicians. My family is full of physicians, and I see even them questioning traditional methods. I would caution against media's rush to blame anything going against the mainstream narrative as "anti-science" or "misinformation." Yes, there are quacks and morons, but let's not put labels on anyone questioning bad outcomes.
Fair comment. You would think medicine would be evidence-based, but a lot of it is pattern matching and working on the 80/20 rule, given the limited time they have with clients.
> You would think medicine would be evidence-based (…)
Your comment sounds like it refers to the front line contacts with the patient.
It has been a while, but my own experience was that (1) the studies I wanted to see did not exist; (2) the doctor was not forthcoming about their own statistics / outcomes; (3) outcomes were not tracked by anyone past (very small N) year; (4) no access to prior complaints against doctor.
I’ll stop the list there, but when things go wrong it is evident that science is not being done.
The best related published account I know of is of the best cystic fibrosis treatment centers in the country. (Sorry, no reference.)
