@mr_bailey's banner p

mr_bailey


				

				

				
0 followers   follows 0 users  
joined 2023 June 14 14:26:38 UTC

				

User ID: 2497

mr_bailey


				
				
				

				
0 followers   follows 0 users   joined 2023 June 14 14:26:38 UTC

					

No bio...


					

User ID: 2497

I think he means the memory is for the coding agent workflow, not the apps generated. I.e. it is for the AI chips, but coding agents (and agents in general) generally require much more memory on the AI server doing inference then chats with users and have grown in usage a huge amount in the last year. The AI coding agents tend to operate with much longer context then your run of the mill chatbot talking with a user (even a chatbot that's doing some thinking). Memory requirements for the KV cache on AI inference server scale linearly with context length, a 500K used context agentic coding session (not hard to hit at all, context length limits for paid plans can reach up to 1m tokens, equivalent to a book 2500 pages long) is going to use about 50 times as much memory in the KV cache on the AI inference server then a chat conversation that maybe (if you're really chatty) reaches 5-10k tokens

The sulfate aerosols are released in the stratosphere, the whole point is that they tend to stay up there as there is very little mixing with air lower down in the atmosphere which is what allows the whole thing to work. The amount that drifts down (as we can see from the natural experiments with volcanic eruptions) would be barely detectable against background levels (<1 point to the AQI scale), if you live near even a modest sized city you already breath much more of the same compounds from diesel exhaust (and you used to breathe a lot, lot more if you were alive before 2010 when the finally switched to ultra-low-sulfur diesel in the US).

High altitude Sulfate aerosols. All serious cost estimates I've seen come in at <$20 billion / deg C / year of warming averted (and the cheapest options are more in the ~$2-3 billion range, depends if you use existing commercial aircraft as a base or develop a new higher altitude aircraft). Combats the warming part of climate change (which also combats the sea level rise part), there are some other aspects of climate change that would not be ameliorated (ocean acidification, plant growth would be harder to predict but is likely a wash, couple percent less sunlight may have a some effect, but we already see a pretty large opposite effect from CO2 fertilization, and the more politically fraught is that some hard to exactly predict changes in weather / drought patterns would be expected), but it would probably avoid a lot of the potential harm. We have high confidence this would work (see ocean shipping, volcanoes (esp. some larger historic eruptions), etc.) without causing a lot of unrelated problems (again see volcanoes). The main risk is some changes in weather/drought patterns (not that we don't see those with the status quo global warming). Even let's us somewhat target the cooling (limited, but we could vary how much cooling we see in the polar regions (where it's more needed/effective) vs. the world as a whole). Plus and minus is the effect is temporary (half life of a couple of years), that means you can adjust pretty quickly, but it also means if you are doing it to a significant amount and suddenly stop for some reason (war with the AIs?) a lot of built up averted warming could happen very quickly.

America not winning doesn't imply that Iran is winning. Both are losing. The winners are anyone still shipping oil (Saudis are making bank, UAE too since they also have a straight bypassing pipeline and were never completely shut out like Kuwait etc., and other oil exporting dependent economies outside the middle east). Parts of the US are winning (frackers, and especially refinery owners, the spreads are huge right now) and larger parts are losing (farmers, fishermen, everyone else).

Some of this seems off to me.

"Diesel and gasoline are still stymied, but, mostly, because Iranian exports are being blockaded by the US." I don't think Iran has ever been a significant exporter of refined products (if anything historically it's been a significant importer do to lack of refining capacity though I think pre-war it was improving in this measure). Damage to refineries in Saudia Arabia / UAE and logistics disruptions or the like would seem more likely?

I'm surprised the oil is flowing so well, a lot can change, but it was only 10 days ago that Saudi Arabia declared Force Mejure on some of it's European contracts and said it wouldn't be able to make October deliveries (some Saudi fed European refineries with longer term contracts are slated to get 0 barrels of their contracted oil delivery in October). This is the kind of thing that's keeping prices (and prices of refined products) high. The Saudi East-West pipeline is still only at 50% after being hit and shut-down a few weeks ago (sounds like repairs are going faster then expected, original reporting was that it might be as much as 6 weeks to get things flowing again, so maybe the Europeans will get some of their Oct. oil after all). This kind of vulnerability though is keeping folks nervous.

I don't think the Iranian government is unified on the issue (indeed this has been part of the problem with the previous MOU etc.). Some parts of the Iranian government want peace now, others don't.

"I dunno, I haven't seen a situation where some Dem candidate is so bad that it actually activates more of their base to come out and cast a vote against them in protest. I bet some local elections it has happened." - Arguably a variation of this has happened in California local elections with Chelsea Boudin and various SF school boards members. You can at least piss off folks enough that they will vote to recall and that at least is some message.

Eventually electric cars will reach a level of adoption where oil is less immediately critical (the demand for ships, trucks, trains and planes which are harder to electrify will be easier to satisfy). But that's a ways off and in the meantime we've avoided crisis by eating into inventories and strategic reserves to the tune of ~6-7 million barrels a day. If thing continue that will eventually have to come out of demand destruction through prices high enough to reduce demand and gas demand is pretty inelastic. I'm not sure you can get 1970's style crisis without the political component, but prices and inflation could definitely be squeezed quite a bit higher.

If things crashed and frontier models and data centers with their chips and hardware were effectively free I can't imagine someone couldn't productively use it at greater then the operating cost. Anthropic is paying $12.5 billion a year for Collosus 1 and best estimates I've seen is their inference margin for it will be >50% so they will sell that for >$25 billion / yr. It takes about $200 million / year for electricity to run Colosus 1.

2008 seems entirely the wrong reference point, the dot com boom would seem a much better analogue. There is clearly a lot of potential for value and the market may get ahead of itself trying to figure out how and where it will be and over invest in infrastructure from FOMO. But the general thesis seems pretty strong (like dot com boom having the internet taking over everything, they were not wrong! yes pets.com went bust, but everyone gets their dog food from chewy now, orders from Amazon watches youtube and door dashes their food, can you even book travel by phone anymore?) and very likely 5 years later it would all be bigger then the peak again. Maybe we'll end up with 'dark' data centers like we had dark fiber for a while, but I would bet not for that long.

The release of ChatGPT was just the wake-up call that something was coming. No one used it back then for more than a trivial amount of real work. People will vary, but I'd put the dividing line for that about 9 months ago. It will take time to diffuse and for processes to adapt. So such conclusions seem premature.

Also, that 2% growth isn't some magic rule of nature, excepting population growth (which is covers about 1/4 of the economic growth over the last 4 years) the default growth is 0% (as it was epsilon close to for much of history) only by investing and inventing things does growth happen. AI is a thing human invented and are investing in, it's quite plausibly the thing that will keep the 2% going another couple of decades. We don't get to do nothing and magically get 2% growth as a baseline and AI is only measured as to what it adds on top of that. Computers, cell phones, internet, cloud services, etc. all of that was apparently worthless as we only grew 2% a year.

A lawsuit would be expensive relative to the value of the goods and they might not win it (the law here is fairly involved, under commercial law it's not clear that he isn't just another creditor against the bankrupt franchisee as other suppliers supplying goods on credit would be). Morally and in the court of public opinion he has a much stronger case.

Also, it seems like there's much better options for the guys? There was an article a while back about a somewhat autistic SF Tech guy that described his situation in which he basically had a subscription service with an Eastern European group that would send him an English speaking, very attractive, attentive Eastern European girlfriend of his type (given the pricing, probably Ukrainian) on a rotating basis every few months (I'm guessing on a tourist visa and they head back when the visa was running out) at an annual cost comparable to that 30k one weekend above. If you're into Asian woman you could prob. get even cheaper with Philippines/South East Asia or the like. This seems a lot more appealing to me if you're already going the escort route.

Would that be the case, but safety people seem to end either end up fixated on very specific, relatively rare crimes (see magazine limits, long guns/assault weapons vs. hand guns), aesthetics (assault weapons) or narratives (suppressors, 3d printed guns) with a limited sense of balance. To the OP, folks using any of those alternates are more likely to kill themselves then anyone else. The low hanging fruit is a pressure cooker and some fireworks (see the Boston Marathon bombing) which any idiot could acquire.

Not to say this isn't a fiasco, but re 4, the US is, on net, energy independent and exports more oil products then it imports. The US refineries are more capable then most and so the US imports a lot of sour crude and exports a lot of light crude as that is a more efficient use of refining capabilities, but that sour oil also predominately comes from Canada (who's oil is largely captive on American pipeline infrastructure), Mexico etc. and likely Venezuela in the future. If the US closed the export of oil you'd likely start to see a larger price divergence between the regional oil markets (more akin to Natural Gas where the US price is substantially lower). Not that this would win the US any friends, but it's something I could see Trump trying. There are a few geographic places in the US that import oil from abroad (often the middle east) like CA and HI, but that's largely down to the Jones Act making it too expensive to ship from other parts of the US.

"Calling Iran's bluff on Hormuz,"

Is it a bluff? The oil is only trickling through. And Iran has proved that in the age of drone warfare there is an attackers advantage going after fixed infrastructure such that even given the vast disparities in capabilities and resources the US and allies was unable to protect the gulf oil infrastructure from Iranian attack.

It's a a good point, number went up during Biden's term, but in real terms it mostly treaded water for us. It has gone up significantly during Trump's second term (not through any particular investing insight, mostly own broad funds plus some extra in oil/resource stocks and tech stocks). We have also 'retired', if involuntarily (both got laid off), which prompted looking at our finances and realizing that under any non-extreme-blackpilled assumptions there is enough there (and in those situations the skills are probably worth more then any extra money), which hadn't really been on the radar given we are in our mid-40s with 3 kids. 'Retired' is in quotes though as within 6 months old colleagues began bugging us to take on some contracting projects which we did to help out folks, stay up to date, make some side money, etc. My wife ended up making more last year doing that then she did working full time and I made about 1/2 so we'll just see how it goes. Congrats on retiring.

But does the spice flow? The previous Iran leadership at least played lip service to working relations with it's Arab neighbors. The new leadership, or fragments thereof, seem much more willing to break norms. The US may be done with the war, but that doesn't mean that the war is done with the US (or at least it's allies/interests). If the flow of oil out of the Persian Gulf remains blocked, heavily impacted, tolled etc. no one is going to view that as a US win, not even Americans if gas prices remain high and downstream inflation picks up as result (both direct from gas prices, and indirect through increased energy costs and imports from directly impacted countries).

The general idea behind the SAVE act seems not unreasonable and in some sense this is the time to do it (unlike in the even fairly recent past it's not clear given the current voting patterns that this particularly favors one party much over the other). That said the politics of it seem difficult to resolve.

Then don't use it, that's the whole point of a free market. Same for the DoW, I have no problem with them not using Anthropic (and from everything I've seen Anthropic was perfectly willing to let DoW unilaterally end the contract early and offered support for a transition period to a new vendor), the retaliation is just excessive, capricious, and completely outside the intent of the law in use (I suppose par for the course for Trump 2.0 admin).

Arguably, though, Anthropic's ethical stance is partly why they are the best, so I'm not sure this is so easily separable. Certainly it has been key to their recruiting and how they have been able to attract the talent they have and they are by far in the lead in model interpretability thanks to this focus. They have also argued in various papers that this approach has led to better model performance (most obviously in false refusals, which will be a thing until we decide we're ok with commercial models teaching people how to make chemical weapons, etc.).

Oh I'm sure that's what they are going for. Though the lesson could as easily be don't even start doing business with the US government which may not be the win they imagine.

This seems to be an entirely different claim though. Is the problem that Anthropic is insisting on certain contract terms around selling it's current products remain in place or that it won't generate a more morally deferential AI? The later seems to be what you object to, but, in theory at least, is not the crux of the current kerfuffle. Developing an AI within a consistent ethical framework is kind of Antrhopic's whole thing and arguably has probably helped them (certainly at minimum, with recruiting). Idk, mileage might vary, but I've found Claude to be pretty nuanced in it's opinions on the use of violence in the context of self-defense and police shootings at least compared to ChatGPT and Gemini which seem to be a lot more proscriptive and it is certainly empathetic to the users position. I'm not at all convinced on your claim. If you're looking at models likely to tell someone off, Grok or some of the chinese models are much more likely. GML 4.7 is too far off the frontier (or alt. to narrowly focused) for me to consider it a strong comparison point. If that's what you want/need/suffices by all means use it, but it's not a replacement (or if it is not sure why DoW is so focused on Anthropic).

A highly relevant aspect is that the government paid Lockheed to develop the F35 under specific contract. It's not exactly commensurate, but would it be a supply chain risk if SpaceX said it was unwilling to launch nukes into space?

That's an inherent risk, they could turn off the tap (to the extent they are able to, I don't believe Anthropic is actually running the hosting) whether they agreed to the contract changes or not. Anthropic offered a 6 month transition period gratis for the DoW to transition to a new vendor so does seem to be operating in good faith.

I mean what will happen is they'll get it from someone else. Otherwise, what exactly are they going to take? This isn't some factory or a warehouse full of inventory. Maybe they can take the current model weights and have something that's obsolete in 6 months. What they want is a Claude 6 that will do whatever they ask it to and for that they need Anthropic and it's employees to cooperate and without North Korea levels of oppression there's only so many levers they have.

I mean, current kerfuffle aside (which you have to admit is highly contingent, there's no way anything like this plays out if Trump isn't president), Anthropic seems to be doing really well commercially? It has the fastest revenue growth of any of the AI companies (and on current trends would overtake OpenAI in the next year or so) and seems to be the leader in integration into workflows etc. Given it's rather paltry free tier adoption and rather high API rates it's likely already significantly profitable on marginal inference basis. I'm not at all convinced that it's ethical stance is hurting it (and it's virtue ethics approach may in fact relate to why it tends to have lower refusal rates then OpenAI and Gemini). I'd be curious on a poll of career soldiers on their opinions on autonomous killing robots (the point of distinction, Anthropic did not prohibit the AI from helping kill people, only doing so completely autonomously), I'd don't think they'd necessarily want to be out of a job.