> I paid ~$800 to rent a 2x H100 SXM node from Lambda for ~95 hours, and ~$400 in OpenAI API fees to generate the Astra trajectory demonstrations.
> a tiny 4B model went from not being able to understand the harness it was wrapped in, to achieving a 1.81x geometric mean speedup and a summed latency decrease of 44.7% across a workload of join-heavy SQL queries
I can’t find it in the article (may have skimmed it too much), but I suspect they didn’t include those ~95 hours in the benchmark numbers.
I think all database vendors know their query optimizers could do much better if they could afford to spend lots of time to derive query plans.
⇒ this may be useful for some workloads, but even then, can you afford to spend hours every now and then to update your 4B model to ensure it still picks a good query plan?
> ⇒ this may be useful for some workloads, but even then, can you afford to spend hours every now and then to update your 4B model to ensure it still picks a good query plan?
I think this would be likely comparable to a scheduled backup, so I think it would be an acceptable maintenance window. However, deterministic algorithms would likely beat re-training (or re-fine-tuning) the model. For example, one could analyze actual distributions or whatever (instead of assuming uniform), and then some plans would automatically be eliminated.
Imo a good thought experiment is to look at places that are hyper-optimized, like compilers. Would LLMs bring anything to the table (architecturally or performance-wise) to a piece of software that has been carefully crafted for decades? (Methinks no.)
The Postgres query planner has had to operate, for those same decades, in a much more realtime-sensitive and restricted environment than compilers. It can only draw its conclusions from summary statistics on tables in isolation, not on their relationships with each other (and even less so when filters are involved). For many cases this is fine! For many others it isn't.
There's a good number of heuristic choices in compilation where, maybe, you could get more optimal outcomes with machine learning - but at the cost of compilation resources, both time and space, and possibly determinism too.
As an example, register allocation is graph colouring, and thus NP complete; a model for producing an allocation plan is learning heuristics that might look at more features in combination than the ones hand-crafted into the compiler. An LLM for the job might do better than a more focused model like a GNN, due to sheer size, the effectiveness of transformers, or magic. But it probably won't do an overall better job than the handcrafted heuristics, because those handcrafted heuristics also tend to compile very, very fast with a small memory footprint, and can be debugged (more) easily when they go wrong.
“Also in September 2023, France and Germany proposed a restructuring of the membership of the EU to allow the UK to rejoin as an associate member, rather than a full member. Associate membership would include membership of the European single market but not the EU customs union, in exchange for participation in the European Court of Justice and making contributions to the EU budget.”
> Just days after a similar (criminal) incident causing a train to derail in France
From what I read, the similarity ends in “both involved trains”. This chain of attacks seems to have used short metal pipes (“Footage published by Dutch public broadcaster NOS on Tuesday showed police carrying what appeared to be metal pipes - about 1.5 metres (5ft) long”) to short a connection, making safety systems signal “there’s a train there”, stopping trains further up the track.
The France incident involved a piece of rails heavy enough to cause derailment (“The rail fragment was estimated to weigh between 250 and 300 kilograms”)
My money is on this being a rogue farmers protest.
The easiest way to hide the incompetence of local european politics, be it expensive diesel or the upcoming economic prices is to blame it all on russia.
Volkswagen is laying off 100k of workers, because they can't make cars people can afford anymore... blame russia! Germans fed up with merz and vote afd? Russia! Macron doesn't even have 20% of support in polls anymore? Russia! Stub your pinky on the corner of the bed? russia!
It seems unlikely they would let any political divisions or legitimate grievances go unexploited. It is their cultural core competency for centuries. That does not mean the greivances being exploited aren't valid, or that the parties themselves don't aid in the fracture so they don't have to worry about competition with parties whose ideology is too close to theirs.
Edit: This is reductive, Russia has tons of culture besides this trait, but it is the fundamental strategy that allowed a landlocked state with no natural borders to survive and expand. When you learn a strategy to survive or cope with adversity, it sticks long after it is beneficial.
But all those political divisions are a result of national politics of those countries. People want higher wages, less taxes, affordable housing, more recently cheaper gas/diesel, and in the last decade (or more in some places), less foreign migrants. If no other political party will do anything about that, people are forced to vote for Alternatives (intentionally capitalized).
Especially after all the anti-russian propaganda being false ("three days of rockets left", etc.), after double standards ("ukraine is being attacked, we must help them!", vs "israel is commiting a genocide, we must continue selling them weapons", and "iran was attacked by our buddies, let's send nato (uk, etc) planes to shoot down their counterattacks"). Americans started a war that hurts us every time we go to a gas station and no politician even points a finger at them, no sanctions, no nothing.
It's not russia exploiting the divisions, it's the eu governments being incompetent to the levels never seen before, be it local national governments or EU levels (eg. 3eur customs on everything from china, chat control, etc.).
Look at this for example: https://cbnunfiltered.com/macrons-approval-rating-falls-to-h... macron has 16% approval rating, while 81% see him "unfavorably"... how shitty of a politician do you have to be to not step down when 81% of the country hates you?
If the governments listened to what people wanted and did at least something in that direction (again, top paragraph), people would not be looking for Alternatives (again, intentionally capitalized). Since they don't the Alternatives are the only thing left.
The way I see it, both things are true - European politicians and major parties have stirred anger by their incompetence, AND Russia has exploited this anger for their own purposes.
EU has huge flaws, but it's still better than Russia. Germans, French and the British may not understand it, but us unfortunate enough to live next to Russia do. There is a reason why Poland, Baltic countries and now even Finland and Sweden joined NATO. Russia is ruled by evil, murderous tyrants (as is always has been), while EU elite is merely lazy, greedy and incompetent.
I'll choose pro-EU parties over pro-Russian parties as long as EU is not bombing children's hospitals and massacring civilians. I can only hope citizens of other countries do the same, or otherwise there's a pretty good possibility my country will face what Ukraine has suffered.
It's unfortunate that so many alternatives to mainstream parties in big EU countries are also more or less pro-Russian. One should be able to vote against the flaws of European Union without voting for closer ties with Putin.
Yes, and Stalin was Georgian, Kaganowitch was Ukrainian (and orchestrated the Ukrainian Holodomor), Breshnev was Ukrainian and so on.
The alternative parties are not pro-Russia. They want to trade with a totalitarian regime just like we trade with Saudi Arabia. And they'll break their election promises anyway just like Meloni and Trump broke theirs.
> macron has 16% approval rating, while 81% see him "unfavorably"... how shitty of a politician do you have to be to not step down when 81% of the country hates you?
When elections come due, following the natural order. Otherwise maybe some bad actor could influence polls - which are easier to manipulate than official elections - to make you resign.
This is a completely all over the place gish gallop. Both can be true: there's real reporting on acts of sabotage that aren't made up, and there are domestic tensions happening independently of that. The skill is keeping both things in your mind at the same time.
All that's being said is, "Russia is known to exploit political conflicts to harm their enemies".
WHY those conflicts exist is irrelevant.
No one is saying it absolutely, unquestionably was Russian interference. Just that it could be. EDIT: OK, at least one person is saying that. But most people are likely thinking, "Well, could be, might not be... dunno."
Why those conflicts exist is actually very relevant when those responsible for them keep using "wawawa russians" as a diversion to avoid doing anything about the issues. If both my local politicians and foreign countries are working against my interests I am going to first vote for better local politicians because at least there I am supposed to have agency.
They certainly seem to be implying it. And not only that, disinformation narratives often to to emphasize domestic political troubles that fan the flames of skepticism against mainstream institutions, and their reply to a topic about Russia was to go straight to that particular well.
When the mainstream institutions are obviously divorced from reality ( Nordstream - "Russians did it themselves", or every mainstream news segment in spring 2022 starting with "unprovoked full scale invasion of Ukraine" where even a casual glance at numbers involved (~ 180000) pales in comparison with what US used in Iraq I or II), you don't need disinformation narratives to elicit skepticism against mainstream institutions. A bit of common sense is entirely sufficient.
One could even argue the mainstream institutions were involved in disinformation narratives (e.g., in the case of 40 beheaded babies or widespread systematic rape on Oct 7th: did not bother to check the facts before spreading it wide).
> Do you claim that Russia doesn't invest in disinformation campaigns against the West?
Maybe it do invest. But that has about zero effect. If Russian disinformation campaigns were working, they would work in Russia as well, but that's not happening.
Old German politicians had many opportunities to say something about the genocide. They decided to support Israel instead and (maybe) this choice played some part in their demise. They miscalculated how angry was the people and basically fired themselves. Should know better after seeing the Democrats fall in the same trap.
When you are surrounded only by people that says "I don't care" every single time, you may forget that those aren't the people that you have to convince that "you care".
The similarity is also timing and geopolitical context. It doesn't have to be an identical method to merit the comparison as the means for sabotage can be variable.
Of course more similarity would be a stronger signal, but we shouldn't miss the forest (rail sabotage) for the trees (shorted circuit).
As far as sabotage goes doing it in a way that is disruptive but causes no permanent damage is a bit odd. As a protest, however, it makes perfect sense.
Economical damage caused by temporary chaos is still damage.
Forcing delays on European trains in special days when a huge number of travelers are expected is the oldest trick in the book. It has been exploited for years and years.
Those aren't the first trains that had problems in Europe recently. Spain has been plagued with thousands of Train delays, cable stolen before a holiday, sabotages and rocks in the railway are also very popular. Russians are obsessed with trains, and the Kremlin even tried to buy the main Spanish train maker disguised as Hungarians, before to be blocked by the Spanish government.
About rogue farmers protest in France... well, exactly the same problem. They are always agitated by far right agents in special moments, and when people scratch a little, this people always are seen taking selfies and money from people associated with Putin. Farmers still didn't understood that are being used every single time, often again their own interests. See the Yellow Jackets movement for example.
I'm surprised people jump to Russia instead of the radical left. Left-wing environmentalist terrorists attached shunts to rail lines to short out the safety signals back in 2020 in the US [0]. Vulkangruppe has been conducting sabotage in Germany, and a 48-year-old named Daniel V was arrested on suspicion of sabotaging Germany's electrical grid because he hated fossil fuels [1]. The French had left-wing environmentalist terrorists sabotaging the rail network in 2024 [2]. The left has been conducting infrastructure attacks all over Europe this decade, but it's probably more politically expedient to blame Russia.
> Vulkangruppe has been conducting sabotage in Germany
It's important to note that AFAIK nothing in this direction has been proven. After the attack, some people claiming to be part of Vulkangruppe claimed they did the attack, but it's all just claims.
Also important to note: the Vulkangruppe's letter mentions JD Vance, but spells the name "Vans" - almost like the authors usually write/read about him in Cyrillic texts.
- Extinction Rebellion and other environmental groups always claim the action pretty quickly and openly in NL.
- If you look at the locations of sabotage, it’s all close to the ‘farmer heartland’. Most left-wing and environmental groups would do it in the randstad.
Most likely it’s farmers. They announced very heavy protests today and also burnt hay near highways to create smoke clouds. Possibly nudged by Russian operatives because the like to sow dissent.
Also, extinction rebellion goes after motorways blocking private cars, not trains which are a sustainable mode of transport they support.
No, I'm pretty sure this would have been a farmer splinter group. The last time they protested with tractors on the motorways it caused a deadly accident, they won't be getting away with that anymore. So they'll be looking for new ways to protest.
Also, today was Prinsjesdag, similar to the state of the Union in the US. It was a day to expect local protests. Nobody in the Netherlands thinks Russia is behind this.
> But I rather a model be suboptimal in encoding that can be used to generate hypothetical reasonable flags in its unpopulated states, than having an optimal encoding which doesn't answer questions or reveal possibilities in its unpopulated states
This encoding already goes too far towards “support the existing flags”.
As a simple example, look at the pattern of the flag of the Comoros (https://en.wikipedia.org/wiki/Flag_of_the_Comoros). Even using just the top colours (FTA: “red, then white, blue, yellow/gold, green, black and orange”), there already are 7 × 6 × 5 × 5 × 5 = 5,250 ways to colour it so that no two neighbouring colour bands are identical. That alone doesn’t fit into 12 bits.
Even limiting one to all colours different, there still are 7 × 6 × 5 × 4 × 3 = 2,520 colourings using the 7 most popular colours. That’s (just) over 11 bits.
And then, you still have to encode the crescent with the four stars.
I think what makes reduce less popular is that it takes two lambdas:
- a slightly awkward one that takes a partial result and the next value to produce a new partial result
- one that maps the final partial result to the result
Also, in many languages, when reading the code, you have to skip initialization of the partial result, read the lambda, and then jump back to make sense of the initial values
I think something like awk’s syntax, with BEGIN and END blocks would improve on that. Example of a first go at such syntax (needs work):
Items.BEGIN
min = ∞
max = -∞
sum = 0
n = 0
ITER
min = Min(min,_)
max = Max(max,_)
n += 1
sum += _
RETURN
average = sum / n
(min, max, average)
Advantages:
- items in the partial results have names, making them easier to understand
- result also is easier to understand
Price paid is wordiness, and you cannot simply write a function name for either of the lambdas.
However, I think the latter only is useful in case the partial result is the final result. There, you can keep
> a tiny 4B model went from not being able to understand the harness it was wrapped in, to achieving a 1.81x geometric mean speedup and a summed latency decrease of 44.7% across a workload of join-heavy SQL queries
I can’t find it in the article (may have skimmed it too much), but I suspect they didn’t include those ~95 hours in the benchmark numbers.
I think all database vendors know their query optimizers could do much better if they could afford to spend lots of time to derive query plans.
⇒ this may be useful for some workloads, but even then, can you afford to spend hours every now and then to update your 4B model to ensure it still picks a good query plan?
reply