MITSardine

Fascinating work and write-up. This is like a vignette of France: weaponized mauvaise-foi, sacred institutions and, more generally, some remnants of an aristocratic hierarchy, and also probably a lot of honest technical incompetence.

Open book need not be easier, it entirely depends on the exam.

This was too long ago for me to remember the details of but, just making an example here, we might have had a course on homogeneization of elliptic problems using a model PDE with a div(A grad u) term throughout, and then the exam was about a PDE with a curl(A curl u) and some other added terms.

Notes carry the method and theoretical tools but you then still need to apply them to a new problem.

So, if anything, our open book exams were harder and a better test of understanding and mastery. You couldn’t rote your way into a good grade.

It's just labor law, these are countries where job security is considered very important. As a result, a lot hinges on having an indefinite duration contract.

The same rule exists in France and you'll have a hard time getting a loan or even a rental if you don't have the sacrosanct CDI (Indefinite Duration Contract), or legal guarantors with one.

It may seem absurd in these edge cases, but it's a way of putting pressure on employers to either hire people for the long term, or let them go.

In this case, it's not necessarily a bad thing in 90% of cases either (because those that don't get permanent employment despite scoring ERCs are certainly very rare), it prevents people from being stringed along in a lab where they're not really building up a career, but rather setting themselves up to be 40 and unemployable, when they're eventually not given the job security the lab would have given 5+ years earlier if they ever intended to/could do so.

You presumably spent 15 or so years in school to learn things of which many only have very indirect applications, and at most a small minority directly applied to your career.

Should we similarly get rid of mandatory education? Most of it is useless after all.

From the article, it sounds more like these funds are the research budget (what you pay other people with). This is quite attractive, students are typically < 200k / PhD, so you can fund quite a few theses with this right off the bat. Basically max out the lab for the grant duration.

I don't know what their salaries would be exactly. This is probably most dependent on where they land, as salaries are very often standardized in Europe. There's usually salary grids per institution dependent on seniority with some milestones being merit-based. Quick google search indicates gross salaries for Professor level (mid/late career) researchers to be around 110-165k€ in NL.

That seems pretty sweet. It's comparable to what US professors make in the hard sciences, as far as I know, with lower CoL than most areas where professors make similar salaries.

And again, salary isn't everything to a researcher. If they can't hire, they're pretty strapped. At this career stage, they're managers, not so much individual contributors. I'd say a maxed out lab for 5 years off the bat is pretty enticing, which also gives time to get up to speed on European funding schemes like ERC grants.

I was a postdoc in the US during Trump's reelection and there were several months where my institution and others had completely cut off scientific staff (such as postdocs, research scientists and engineers) recruitment due to NSF defunding and other threats. Even now, they got taxed on endowment and lost basically 10% budget. This is considerable, and a source of stress for researchers and their current/prospective staff. You can't work properly if you're under the Damocles sword of being laid off / having to lay off your staff.

Stuff just gets done, I guess? Projects move faster, people onboard faster with less intervention, etc. The speedup seems noticeable enough that it doesn’t need precise measuring.

As a disclaimer I haven't read the article, nor do I know much about simulating instruments in particular, but I just wanted to point out that accurately simulating the physics of a musical instrument is most likely still a very difficult problem.

I have no doubt there's been analytical/semi-analytical models around for decades. I mean a program that can take an arbitrary geometry or class thereof with specific materials and simulate the high frequency vibrations and model interactions with the body with high fidelity (not through ad-hoc models) is probably still out of scope of real time simulation.

My point is really that there's often families of models that deal with one thing, from semi-analytical first coded in Fortran in the 80s that can run in milliseconds but is only valid in certain configurations with a low degree of accuracy, to "first principles" simulations that may well require a supercomputer to produce results to a useful degree of accuracy (and not in real time). So, just because you see someone claim they can "simulate X", and then another makes the same claim 40 years later, that doesn't mean they're doing the same thing.

For instance, aeronautics has XFOIL. It's a semi-analytical model first devised in the 80s that computes aeronautics coefficients for a certain class of airfoils (NACA). My understanding is it's a very clever, and industrially significant, piece of code, but ultimately it works in a narrow regime with some heavy simplifications. You can now get results from this in real time on a webpage. A proper CFD calculation to a NACA wing will take in the order of minutes to hours on a workstation (depending on requested precision and settings, e.g. speed of air), and while closer to first principles, it's still using physical simplifications (RANS). So yeah, although nominally people have been "simulating airfoils" for 40 years, the techniques have refined considerably, and will continue to do so (practical LES and, someday, DNS). It might be another century that people are still "simulating airfoils" in ever more accurate (nailing down within the constraints), high fidelity (lifting constraints) and generic ways.

Back to instruments, this is a difficult coupled problem, in fairly high frequencies (high frequencies = more expensive), with possible fluid-structure interactions, not to mention the geometries are fairly complex (to even get a workable mesh to begin with). My uneducated guess is we're still at either semi-analytical, or at the "considerably simplified first principles" stage for this type of problems. Just like DNS, I'm sure you could "just resolve the scales and run it through a simulation with a really tiny time step", and this is liable to be similarly expensive as DNS (million dollar single simulation). Additionally, they have to deal with the human ear, which is perhaps more unforgiving than an error plot on drag or lift. So I wouldn't dismiss news of instrument simulation as stale just because someone made something that produced similar artifacts in the past, as the methods will continue to evolve considerably.

Great article, thanks for sharing.

I didn't know (but should have assumed) AI-generated podcasts existed. That's depressing.

I imagined if mankind had the ideal machine, that could automate anything, we would get rid of dull office work and back breaking physical labor, but not the things that are actually enjoyable: sharing with each other, entertaining each other, making art. I imagined a lively world of live performance and creation, since all subsistence work had been taken care of. Instead we might end up in the world of fifteen million merits.

It seems people don't mind letting their minds be hacked by machines that can create the form of what they find enjoyable, if not the substance. But I guess there's always been slop and the public for it. To imagine actual people wasting their limited time on Earth listening to these GPT logorrhea podcasts is truly depressing. The unchemical soma.

What are we even supposed to spend our days doing in this bright future of the AI champions'? Stop automating away the things that give people purpose, tackle real problems instead.

To take this further, don't LLMs justify lowering the "barrier to attention"; i.e., if it only takes Claude's and not the hacker's eyeballs on the software, won't people find vulnerabilities in custom software for one too?

Besides that, one could easily imagine software created for similar purposes ("make me a file editor") by the same tool or handful thereof (claude and a very small "etc" for completeness) might share similar vulnerabilities, so this kind of broad net might be even cheaper to cast than one might imagine at first.

What are these languages? I'm curious because C++ is the standard in my community so it's all I ever hear about (and Rust from the adventurous few).

In French, there's a saying: "la nuit porte conseil". Roughly translates to "the night advises", and it means it might be better to sleep on it.

I recall my father (also a mathematician, incidentally) often repeating this to me.

Not an attorney so I don’t know the details but it is definitely possible to leave the US several months in a row on a J1. When I did it (also as postdoc), it was an involved process that escalated to the (vice?)president’s office to get permission so there are clearly questions the university needed to address... I didn’t have any visibility into it all but what I was told about regarded taxes mainly (since getting paid abroad).

So technically possible but also a tall ask (I didn’t know at the time of asking and my PI went with it).

I then came back and carried on without any immigration issues.

If your country is Italy that might be the case, but groceries are at most 30% more expensive than France, and some are nearly the same price (vegetables). Meat and fish do cost an arm and a leg (100% tax on border crossing).

Meanwhile, median net salary in CH is 5'000-5'500 per month, double to triple its neighbors. So food is actually very affordable.

The food that costs more is the one someone cooked for you, which is logical considering the cook is likely paid more than your engineer (assuming that's your case) salary. But then again, minimum wage Italians are not eating out at the restaurant with any frequency. If you were an engineer in Switzerland instead, you could afford eating out there. The restaurants and terraces are never empty, anyways.

Now, if you want to enjoy a beer in the sun, you can get a 2CHF can at the supermarket and go fire up a barbecue at the lake of Zurich, I see people doing that all the time.

But 1% is on the total held, not on the capital gains, right?

If that's the case, it affects earnings quite a bit. Say your investments beat inflation by 3 percentage points, you're effectively down to 2 percentage points after tax, so a 33% reduction in income.

You'd have to break this down into "archetypes" because like 90% of what you pay for is specific: - retirement - healthcare: negligible expense for most young people (for instance, about 80% goes to the 65+ in France) - unemployment (a little more universal, with some large variations still)

Then there's everything to do with children (from direct subsidies to public schools or kindergarten slots), education (not everyone goes to university, for instance), and other subsidies that are income-dependent (two common ones in France are rent subsidy and a salary top-up for low-but-not-too-low incomes).

Plus, ultimately, 100% of what comes in goes out (modulo administrative costs) at the global scale, so you can't just average this or everywhere looks the same.

This is something I could almost never be bothered to do before, but I can now very lazily set up large parameter sweeps and visualization scripts to really probe things. There's a danger of "analysis paralysis" but I've still found it quite useful. Although I'm not sure it saves me time as much as sanity.

On the other hand, if it can detract PhD factory foremen from hiring students for the sake of publishing, and let them focus on advising fewer students better, I can't say I see the harm in that.

On the contrary, my experience of US academia has been that (graduate) students are very much students, who take a lot of classes, are graded seriously with the possibility of failing, are mentored rigorously (the author even says "the classic one hour a week meeting", which I also witnessed there), and in fact enroll in a program more than they are hired directly.

I did my PhD in France where we were legally employees like any other and did 100% research with like 100 hours training over the three years which could be 5min MOOCs counting for hours or classes the professors would sign us off on. We were hired by a specific researcher for a specific topic, unlike US students who join a broader program and explore their own directions more. My mentoring was drinking coffee with my advisor and colleagues and the odd e-mail exchange the day before turning in a paper.

I believe Germany and quite a few other European countries are similar. Any country that does 3 years PhDs is bound to cut on the student part of things.

In the classic division, "teaching" consists in giving undergraduate classes, and "research" consists in the whole spectrum between working all on your own and managing a PhD factory (3+ students a year).

So this article is really not saying anything controversial in the strictly ontological side of things, in fact it's already a relatively common stance to prefer supervising few (or, more rarely, none at all) students.

This researcher is saying "when I consider hiring someone as a workhorse, I might prefer AI instead"; what's the harm in that? Too many PhD students are used as disposable cheap labor, seeing little personal growth in their PhD journey and being generally neglected and abused.

It's an interesting paradox: the more we made computing accessible, the less we got out of it.

When a PC was expected to boot to an OS and not much else, we had all the freedom - by necessity - to tinker and learn. Hardware was barely enough for most day-to-day usage, so we upgraded relatively frequently and got to know the physical innards as well.

This is all so streamlined today that even computers can be smartphones with "apps", or even just a browser that gets you to google slides and everything else (or the MS equivalents). It was probably a necessity that, as computers became infrastructure, they would become simplified, so 90% of the population can indeed file their tax return online (and the remaining 10% have their younger family members do it).

This also means that people nowadays simply don't know that they can walk into any second hand store and get a $200 PC with a warranty that'll be much more productive than any smartphone if they have the knowledge to use it properly. But was there really a loss? These are, for the most part, people that would not have been able to hop on the internet wagon if it'd relied on maintaining a linux distro at all. That's regarding adults; children now do indeed grow up with walled systems for the most part, and that might be a loss.

The build quality and usability on mac laptops is something else, I've yet to see even 2k€+ laptops that people typically get for their jobs that aren't a pain to use without a mouse and monitor. Whereas I'm sitting here in front of my macbook and not touching the mouse next to it most of the time.

That's definitely valuable, but not for a child in my opinion, it's the type of luxury equivalent to a Mercedes over a Renault. Perfectly defensible but, just like a Mercedes is hardly a starter car, I don't think an MBP is that fit for a starter PC. It's also mostly useless if you're not traveling for work regularly.

That said, does any of that even matter any more? People were learning Blender, programming and whatever else 15 years ago on low to mid range machines already. The equivalently priced - or dirt cheap second hand - machines of today are multiple times more capable at everything. Stick Linux and a $5 mouse in it and you're 90% of the way to a macbook pro in terms of user experience.

That's to say, I agree with the core of the article: kids will make the most out of the least. But I disagree that this particular laptop is a necessity or a boon for that. If anything, it's a hindrance for being a mac.

In an organization, the number of sequential steps doesn't really scale with number of participants, does it? Rather with dependent steps of the tackled process; say, devise building permit request, await approval, purchase materials, move materials to site, hire workforce, etc.

Theoretically, each of those steps is parallelizable to some extent. Amdahl's law equivalent here would be that some delays are outside the reach of an organization to improve. For instance, a building permit will take the time it takes to be examined based on an external public administration.

I'm probably missing most of your point, but wouldn't the fact that we have inverse problems being applied in real-world situations somewhat contradict your qualms? In those cases too, we have to deal with noisy real-world information.

I'll admit I'm not very familiar with that type of work - I'm in the forward solve business - but if assumptions are made on the sensor noise distribution, couldn't those be inferred by more generic models? I realize I'm talking about adding a loop on top of an inverse problem loop, which is two steps away (just stuffing a forward solve in a loop is already not very common due to cost and engineering difficulty).

Or better yet, one could probably "primal-adjoint" this and just solve at once for physical parameters and noise model, too. They're but two differentiable things in the way of a loss function.

> You could in principle create a simulation with the same mathematical properties as the physical world but no one has ever done that. I'm not sure if we even know how.

What do you mean by that? Simulating physics is a rich field, which incidentally was one of the main drivers of parallel/super computing before AI came along.

As I recall, a lot goes into DF's world generation, including erosion simulation and the like to achieve a realistic result. That game is the embodiment of over-engineering, after all.

That's pretty impressive! However this workflow could have trouble dealing with the types of meshes coming out of those 3D generative algorithms. Geometries are arbitrary (not simple geometric shapes) so you'll have to fit with NURBS, and the meshes are noisy so that'll struggle/be somewhat arbitrary (what do you consider feature, what do you consider noise?).

However you highlight what I think is the way forward: using scriptable CAD that can leverage LLMs or, maybe in the future, specialized generative algorithms that output in a sane geometry specification.

Some of the defects are attributable to the critical:

> AI models generate meshes using "isosurface extraction" or similar volume-to-mesh techniques

This creates the "lumpiness", the inability to capture sharp or flat features, and the over-refinement. Noisy surface is also harder to clean up. How do you define what's a feature and what's noise when there's no ground truth beyond the mesh itself?

Implicit surface methods are expensive (versus if-everything-goes-right of the parametric alternative), but they have the advantage of being robust and simple to implement with much fewer moving parts. So it's a pragmatic choice, why not.

3D generative algorithms might become much better once they can rely on parametric surfaces. Then you can do things like symmetry, flatness, curvature that makes sense, much more naturally. And the mesh generation on top will produce very clean meshes, if it succeeds. That is a crucial missing piece: CAD to mesh is hardly robust with human-generated CAD, so I can't imagine what it'd be with AI-generated CAD. An interesting challenge to be sure.

Yes, what's preventing the LLM from running myCommand > /tmp/out_someHash.txt ; tail out_someHash.txt and then greping or tailing around /tmp/out_someHash.txt on failure?

I'm only discovering Literate Programming today, but you seem very familiar so I might as well ask: what is the fundamental difference with abundant comments? Is it the linearity of it? I mean documentation type comments at the top of routines or at "checkpoints".

I'm particularly intrigued by your mention of keeping old code around. This is something I haven't found a solution for using git yet; I don't want to pollute the monorepo with "routine_old()"s but, at the same time, I'd like to keep track of why things changed (could be a benchmark).

With LLMs this fast, you could imagine using them as any old function in programs.