academia_hack

Fair. My point is not that all of Marxism is correct (or even most of it) but rather that the specific pattern he and other theorists in that school of thought observed when developing the theory of historical materialism seems to strongly match a lot of what's playing out in AI.

AI, doesn't feel all that different to the mechanical loom that was the impetus behind a lot of Marx's early critiques of industrial technology.

You've got a really cool invention that replaces what was previously a skilled artisan trade. The wealthy then use this invention to excoriate the artisans and render the loom operators as replaceable commodities.

Dario is essentially writing a blog post about why only he (and a few other approved elite companies) should be trusted to own the loom and then make it available for everyone else to labor at. Like the capitalists of Marx's era, he genuinely believes it is better for everyone to have this arrangement and that it is natural people like him should sit as benevolent gods at the top of the pyramid.

It's fascinating to me to see the same underlying mechanisms that gave birth to some of the greatest failed political and social experiments of the 20th century bubbling up again so clearly with AI today.

I'm not terribly optimistic. I don't think humans have really figured out what to do about historical materialism. More just remarking that this is a blog post that almost literally reads "fellow elites, let us ensure only we control the means of production. It is our duty to humanity to control the masses."

I think you're reading too much into what I'm saying. We've got a whole legal system developed over the course of thousands of years which is well suited towards regulating the effects of "intelligence" by regulating what people _do_ with it.

Lobbying for a completely parallel regulatory framework to restrict who is allowed to run specific types of general purpose computer programs, just because we can imagine that some of those "general purposes" might be bad, feels a lot more like an attempt by capitalists to carve out a special kingdom over this specific technology for themselves.

For example, I'd be 100% in favor of holding anthropic criminally liable any time their AI enables someone to commit a crime. But I'm opposed to banning the export of GPUs above a certain size just because Dario doesn't trust what someone like me would do with them if he couldn't snoop on my conversations.

Dario's proposed approach is a classic example of capital attempting to control technological advancement and the means of production. For the first time in human history, any member of the working class can just about afford to have a team of expert scientist/physician/lawyer/engineers working directly for them.

Super intelligence (the ability to have many smarter minds than your own reporting to you) has always been available to the wealthy and powerful. They used it to invent nuclear weapons, cause climate change, mechanize warfare, and pillage the global south.

Now that this same tool is on the cusp of being available to everyone, capital is starting to panic and throw up fences. They want to pump the breaks on a revolution they know they can't control. They see what they've done with super intelligence and (perhaps not unreasonably) fear what the masses will do with that same power.

The essay includes specific regulatory measures to prevent developing countries from accessing hardware and software needed to benefit from technological development. The US government is constantly applying new perversions of arms control regulations based on the preposterous claim that an LLM is somehow an "arms" in order to restrict Chinese access to GPUs. They are willing to prevent virtually any country on earth from hosting and developing AI models just on the slight chance those countries resell or share AI knowledge with China.

This article explicitly asks for the US government to step in and push for further restrictions on model distillation, access to open frontier weights, and access to hardware. It also pushes heavily for independent auditors to monitor and control "AI companies" and for greater observation and control over of what people use AI models for and who provides them.

I'm tired of tech billionaires lobbying the US government to make an AI patriot act that gives them unprecedented control over speech, trade, and technology. The narrative is grotesquely transparent:

1) AI is a dangerous technology that can literally end the world.

2) Only me and a handful of other [people like me] should be trusted to determine who can use it and how.

3) The state must use its coercive power to support my control of this technology to the exclusion of [people not like me].

Reading the definitions here it looks like France is basically banning 15 year olds from using interactive websites. Platforms like Steam, Xbox Live, Minecraft, the New York Times comment section, etc all fall under the super-vague definition of social media services here: https://www.conseil-etat.fr/avis-consultatifs/derniers-avis-...

It's inordinately difficult and expensive to start an LLC or SA in some EU countries. It's even difficult and expensive to _stop_ an LLC and dissolve it. Huge amount of risk and cost on founders and a huge distraction from running a business.

I think that EU-Inc _could_ be an improvement, but it needs to avoid the committee laundry list of ideas/requirements/form fields that plagues the EU startup ecosystem. My worry is that the end result will require notarized declarations of honour, financial plans stretching decades into the future, 30 page business plan documents, reams of corporate governance documents, and tons of other nonsense to protect against the perceived risk that someone who failed at starting a business once fails a second time.

There needs to be UX requirements on the process from day one against which the end result is judged. (E.g. "a company should be able to register in x days", "a complete application should be no longer than y pages", "application costs should be less than z euros").

++

Anecdotally, I find you can tell if someone worked at a big AI provider or a small AI startup by proposing an AI project like this:

" First we'll train a custom trillion parameter LLM for HTML generation. Then we'll use it to render our homepage to our 10 million daily visitors. "

The startup people will be like "this is a bad idea because you don't have enough GPUs for training that LLM" and the AI lab folks will be like "How do you intend to scale inference if you're not Google?"

I hear this a lot (that children learn languages faster, or the corollary from various app ads that the best way to learn a language is to do so like a baby does), but is it actually true?

It takes children a very very long time to learn a language and they're quite bad at it for many years. I've even met some teens/young adults who are only borderline literate in their native language after years of schooling and immersion.

Everyone remotely competent in AI in the federal government that I know has quit in disgust over the past 6 months. I know zero talented AI people who are looking to take a cut in pay, benefits, and career stability to sign up for a new job working for this administration.

As a result, there's zero chance even the sensible parts of this strategy won't just end up coopted into multi-billion dollar Palantir contracts to deliver outdated llama models behind some clunky UI with the word "ontology" plastered on every button.

At least to me, loyalty _is_ the benefit. I can't conscience working for someone I hate or someone who I don't feel like I want to help succeed. I've definitely quit jobs before just because the senior leader in my reporting chain was replaced with some smarmy windbag I didn't believe in.

That's not to say it's _much_ of a benefit, but if the only thing a job gives me is a market-rational amount of dollars and health benefits in exchange for life-hours, the invisible hand ensures I can find that virtually anywhere.

Do you think there is a role for the federal government to employ engineers for any purpose? 18F was a general purpose engineering group that built software for many agencies instead of having those agencies pay vast amounts of money to Deloitte/Accenture/Booz Allen etc for a worse quality product.

Check out: https://trafficcamarchive.com/ for an example.

The data collection isn't even quiet. There's an entire cottage industry of companies that scrape these traffic cam feeds, store everything for x numbers of months in low-cost cloud vaults (e.g. glacier) and then offer lawyers/clients in traffic disputes access to footage that may have captured an accident for exorbitant rates. It's a remarkable little ecosystem of privatized mass surveillance.

True that! I use probably 15 days of "unlimited" leave and still manage to feel guilty about it.

The frustrating thing for people in fed jobs is that if you hit your 13 days that's it (during your first 3 years in government). It can be impossible to get PTO until you build up hours again. You have to either quit, negotiate LWOP (often seen as a performance adverse metric on your record), or work. So if you land a sweet concert ticket, see a flight deal, have a friend get married, etc. you better hope you've banked up the leave for it. Since you gain hours every 2 weeks (4, 6 or 8 depending on service) you also start out in government with virtually no leave and can't actually take a 2 week trip until you've been there almost a full year.

USDS is great! I know people who have made a huge impact there and if I personally were to go into government from tech it's where I'd look. They are situated at the White House which allows them to be hired at a higher level than normal federal jobs (up to GS15, though still lower than comparable private sector work) and then they get sent out to various agencies by the White House to try and fix things. In practice though, USDS is a tiny tiny drop in the bucket compared to what federal agencies actually need. Maybe if every agency had a digital service of their own the model could work.

The federal government is an enterprise with 4 million employees (more than half in DoD as military or civilian). So the handful of people at USDS are basically only sufficient to swoop to fix the most dire of dumpster fires.

In terms of benefits, here's an anecdotal comparison with a senior engineer (5-10 years experience) at a mid-level start up I worked at.

* Federal Pay (GS-12): $100,000 * Startup Pay: $150 base + $25 k bonus + equity

* Federal Health Insurance (United mid-tier plan, no family): $2,500/year * Startup Insurance (United mid-tier plan, no family): $0/year

* Federal Leave: 20 days (after 4 years in federal government) * Startup Leave: Unlimited

* Federal Sick Leave: 13 days * Startup Sick Leave: Unlimited

The pension I'm talking about actually isn't the TSP (which is fine, but slightly more expensive than comparable Vanguard funds).

All federal employees must contribute 4.4% of their salary to the FERS now which is taken out of their base pay just like their health/dental/fegli. It used to be 0.8% but congress gutted it a few years ago.

FERS takes decades before it's more than pocket change and the same money invested in the market would yield higher expected returns without requiring you to work 20 years in gov to benefit from it.

Counterintuitively, scaling government down goes hand in hand with increasing the attractiveness of the civil service.

Right now if a government agency wants to do something like make a webform where you can apply for a passport, they have zero web developers on staff who can do it. Instead they must pay a team of non-technical officials and lawyers to make and adjudicate an RFP. Then pay a contracting firm to put a developer behind a government computer to do the actual work. Putting this contractor in a seat can easily cost the taxpayer $500k a year despite the contractor only receiving $130k of that money. The rest goes to the HR department, IT Department, C-Suite, lawyers, lobbyists, and shareholders at the contracting firm. The government has their own HR/Lawyers/IT too, but the contractor can't use those so the tax payer ends up double-paying overhead and missing out on economies of scale on every contract.

This is one of the many reasons government websites are always $50 million dollar boondoggles that an intern could have done better. The government ends up spending millions of dollars feeding leeching middle-men before they can hand that money to a mediocre dev deep in the bowels of Accenture's cheapest subcontractor.

If an agency just could hire a few strong web developers directly and then assign them to whatever task is needed during a particular sprint, we'd see a massive reduction in cost and increase in the quality of engineers working on our country's most important work. But most agencies are literally not allowed to spend more than $120k on an in-house engineer, while no one bats an eye on them spending 5 times that on an Accenture contract placement.

Totally. I think comp is a necessary but not sufficient precondition for fixing government technology. The actual solutions (good authentication and least privilege systems, robust monitoring, rapid intrusion detection and response, secure by default system architectures) all take talented people to execute and the government doesn't have enough of those in-house. Instead most systems are built with a 7-figure contract to Booz Allen and friends and then maintenance and sustainment is left as an exercise to the reader.

Until the US federal government pays civilian tech talent competitively, this is always going to be an issue.

Your typical hands-on-keyboard blue team engineer in federal government is a GS-12 getting paid around $68,000 per year (or $99k in very high cost of living areas like DC). They have expensive health benefits, 13 days of PTO a year, put a huge chunk of their paycheck (almost 5%) into a mandatory pension plan that consistently underperforms the market, and can literally go to jail for making mistakes at work depending on the statutory context they work in.

The best people in these jobs burn out fast and quit or they end up having to abandon IC work for GS-14/15 jobs (max pay is around $190 for those) in order to keep up with cost-of-living and justify their careers.

As a result, you have almost zero genuinely capable principal/senior engineers in government who have the authority to architect complex IT systems for security. Instead you get contractors who charge the taxpayers enormous overhead costs and cut corners wherever possible.

If there's one letter to write your congress person to improve government - my vote would be for civil service reform to attract and retain actual top tech talent. They've done it for doctors and lawyers (both of whom can get paid well above the $190k GS pay ceiling), but engineering is still not treated as a comparably skilled professional trade.

I've giving "Accept - Minor Revisions" to every paper I've peer reviewed since getting my PhD other than two that were outright plagiarism. Figure it's important to the morale of grad students to get some positive validation and the vast majority of published research is garbage anyways so I don't feel particularly inclined to defend the trash heap as an unpaid reviewer. In practice, I find that I've tipped the scales in favor of a lot of borderline papers over the years and am quite happy about that.

Dan's fantastic. Been interviewed by him in the past and was super impressed by his interest in understanding technical nuance and communicating the story accurately to a lay audience. He's definitely not a headline chaser like a lot of folks I've met in tech journalism. Ars is lucky to have him.

They don't even copy and paste the text! At least in my experience, once done with peer-review I spend a few hours wrestling with the "print ready" LaTeX template provided by the publisher to get my paper and tables to render correctly. Often they'll have an out of date LaTeX compiler or something else in the pipeline, so there's actually quite a lot of unpaid labor involved in getting a paper into a PDF that a journal will host.

The main services they supply are basically document hosting and search functionality. Both at extortionate rates.

Honestly they were just there in the Wikipedia text I copied. I've seen them in more modern texts to help with pronunciations and translation (if I recall correctly some words have different meanings depending on the length of the final vowel but that can normally be determined from context). Romans sometimes used the apex to denote long vowels which would have otherwise been ambiguous but I think it wasn't as commonplace as in textbooks today.

Exactly this! It gets even cooler in this example too because the meter for "Odi et amo" elided to "Od'et amo" directly parallels the scansion for "excrucior" (long syllable, short syllable, short syllable, long syllable). So the two concepts that start and end the poem (love+hate, and torture) are also linked by how they are pronounced. Incidentally, that linkage is also the message of the poem itself.

These two lines are basically just Catulus' being a complete show-off. And IMO, some of Ovid's work makes Catulus look like a bit of an amateur by comparison.

Classical latin poetry is like 10% being able to write down clever ideas and 90% showing off your grasp of grammar and vocabulary such that you can pose and solve incredibly difficult linguistic puzzles. I think Sanskrit is pretty similar in this respect too.

I'm really sad at how much Latin I've managed to lose since my school days. It's really an incredible language and this stack exchange post shows some of that versatility.

Because the words in Latin contain dense grammatical information in their spelling, you can be much more flexible with word order.

This gives classical poets the ability to do crazy things with word ordering to create "word pictures" where the structuring ordering of the words conveys some additional meaning. This can be done in English too, but classical Latin is almost made for it.

For example, Catulus 85:

"Ōdī et amō. Quārē id faciam fortasse requīris.

Nesciŏ, sed fierī sentiō et excrucior."

The translation Wikipedia gives is: "I hate and I love. Why I do this, perhaps you ask.

I know not, but I feel it happening and I am tortured."

But there is so much brilliance in the structure of the poem that translation cannot really encapsulate. The last word "excrucior" (I am crucified) references a relationship between the structure of the first and second line. Each verb on the first line has a "mate" on the second. For example: odi (I hate)<->excrucior (I am tortured), requires (you ask) <-> nescio (I know). If you draw lines connecting these mates to each other, they form a number of crosses - referencing the "crux" in "excrucior". The poem literally depicts the torture instrument that is Catulus' love.

Even more remarkably, this poem follows a strict metrical standard dictating the order of long and short syllables: https://en.m.wikipedia.org/wiki/Elegiac_couplet and it achieves this meter in part due to the use of elision in the opening of the poem, where two vowel sounds get merged due to the ordering of words. "Odi et Amo" is read as "Odet Amo" as the the love and hate crush together and evoke that sense of pressure and torment that underlies the couplet.

Classical Latin had so much capacity for structural complexity that is really remarkable. It's not just that you can say more stuff with less words, but that the allocation of information in the grammar allows for entirely different expressions than you could make if word order dictated meaning.

It's a pity the state of the art here is still so dire for basic use cases. Things like managing group access to files in directories, or giving people field-level permissions to specific columns on specific database records.

All of these languages (OPA, Cedar) are nifty but assume you have a whole team of devops engineers dedicated to replicating slices of your database state and getting them to the policy engine at the right time without discrepancies. Likewise, systems like Permit.io, OPAL, Topaz and SpiceDB claim to solve this problem, but they do it by adding an extra API call to everything your application does and then potentially having a permissions database which is in a divergent state from your real application.

I was really hopeful about Oso for this since it is a policy engine that operates over the applications data as a library (rather than a service) but the development on the actual library appears basically abandoned in favor of their SaaS and it's currently full of weird quirks, bugs and outdated dependencies - especially around the filtering tasks (e.g. show me all the blog posts a user is allowed to read). I'd happily pay for a Oso library that worked, but it seems like they've hyper-focused on SaaS.

Really would love to see one of these languages solve the data access layer for regular applications, everyone immediately seems to run to kubernetes at Google scale when I'd kill for a nice way to write permissions that can sit on the same server as my application and do the job reasonably well.

As an atheist, I've been thinking a lot about this in the context of the factors that made me atheist in a way that would've been not just repugnant, but downright unimaginable, to the previous 20+ generations of my family.

I think the individual rational case for abandoning religious institutions is really compelling. Goodness without gods is obviously possible, and villainy with them as well. The religious machine consumes time and resources and the benefits it yields are ill-defined and diluted by its many harms.

However, theology and religion represent the state-of-the-art of a millennia long effort to establish meaning in the human condition.

Today, many academic disciplines look down upon theology when it was, until very recently, the core of what it meant to be a scientist or thinker - especially in Christian/Muslim/Jewish traditions. It's crazy how many in the social sciences, much less the physical sciences, have never engaged with the religious canons that underpin their disciplines. Not so much the bible stories, as the exegesis that gave birth to academia.

Even for the average person, the Church served as inexpensive frontline mental health care (esp. for bereavement, anxiety, and guilt), a social safety net, a source of community, and a source of simple, generally reasonable, heuristics to evaluate daily decisions. These are all things my "enlightened" post-religious life struggles to provide me.

It's totally fine for society to evolve away from religion. I think faith is in inevitable peril as our understanding of the world grows. However, we may be throwing out the proverbial baby with the bathwater by making the assumption that it is possible to simply skip out on something that sat at the core of most human lives for millennia with no ill effect.

My personal take on the land of cyber security frameworks - and especially security standards - is that a good security team should be able to read through a list of controls (e.g. those in NIST 800-171) and express a reasoned opinion on each one with respect to the company's security posture. They are fantastic tools for reminding you what things you might have overlooked and driving a discussion about how your organization is approaching security - basically regardless of what type of company you run.

That's where the value stops - once you give lawyers, policymakers, and insurance companies access to these documents it becomes an unending game of regulatory capture, responsibility derogation, and box-ticking.

You end up with people who have zero context for technology running around demanding to see evidence that your smart toaster implements 12.2.14.1.5b "The centralized time server must enforce separation of duties" before it can be added to the network or some other such incoherent nonsense.

These standards always start in the right place, but they get used in the most frustrating ways because people who don't understand how technology works are, invariably, the auditors and assessors who apply these standards since true technologists can easily find more gratifying jobs doing literally anything else.