I was reverse engineering a medical device, and had to do a lot of trickery to get Opus 4.5 - not even Fable/Mythos, Opus - not to trip up its fucking CBRN filter.
What happened with Fable is basically what I feared when they announced those restrictions. They took the shitty Opus CBRN filter and made it even worse.
I pity the fools trying to use Anthropic AIs for anything biotech.
Recent image models are advancing rapidly at prompt adherence specifically, and being able to iterate on the same image is propelling them even further. Images 2.0 being the poster child of this "agentic iterative image composition" approach.
Images 2.0 isn't anywhere close to the kind of detail control I'm talking about.
It's the opposite of a skill issue. No image generator is anywhere near the ballpark of pro-level manual Photoshop or Illustrator editing for individual elements in an image.
If you don't understand this, try precisely kerning the text in a generated book cover to handle letter combinations like A and V.
This is one of the big problems with GenAI. You can do new things with it, but it's crude Dunning Kruger good-enough-if-you-don't-ask-for-more creativity.
The pros can see what most people can't, and the flaws and missing features are frustrating and obvious creatively, not just in terms of production values.
We went from "AI can't generate text that isn't at least 20% typos and it always looks like shit" to "some letter combinations aren't kerned to perfection sometimes and adjusting that with prompts is hard". In a couple of generations.
It's RLVR tuned, but not to the ChatGPT level of brain damage, and it's still backed by a fuck off huge pool of model weights - which matters for what you call "touchy feely stuff".
The message isn't subtle, and isn't meant to be: "we don't care how, but we expect you to stick your nose into AI tools and find some way to fit them into your workflow".
Which indicates: the management believes there are productivity gains from AI use, but adoption lags due to inertia and reluctance to change existing workflows.
> Which indicates: the management believes there are productivity gains from AI use, but adoption lags due to inertia and reluctance to change existing workflows.
Methinks adoption lags due to management's inability to align incentives such that productivity gains are rewarded.
You say that as if "align incentives such that productivity gains are rewarded" isn't one of the hardest, most fundamental problems in all of organization and management.
It also indicates several different levels of 'cannot manage their way outside a paper bag'. If you had a construction foreman who decided because nailguns would be the way of the future vs hammers, therefore the metrics would be based on the number of nails used and wound up with thoroughly nailed pieces of lumber but no houses built, he belongs fired as a complete incompetent who has absolutely no business on a job site.
Due diligence, judgement, and ability to know what the hell is going on are essential skills for management. The token metrics are a complete abdication of all of the above. It isn't a cream you just slather on to boost productivity.
more of "we whined and cried and screamed that we needed new budget in order to buy these tools or we would literally die. now we have them, they don't work as well as we hoped, they aren't leading to productivity gains, and they're actively alienating our workforce and users alike. we're so screwed we literally have no idea how to do reverse this."
ChatGPT alone is among the most popular apps ever made, and it's available both inside and outside Apple's walled garden. Letting it reach audience in countries where Apple doesn't have much of a foothold.
I do wonder if new Siri is any good though. Apple used to be a genuine AI leader, but they totally sleepwalked through LLM revolution, and Siri's response quality was a sad joke for a while now. Did they bring it up to modern standard?
I don't think so, i don't think they want to be in the LLM laboratory business. They just want to leverage the technology to make money not invent it. Hence the reason why they made a deal with Google to license Gemini, let OpenAI and Anthropic fight it out while Apple just keeps making sales. I think they're betting that in the long run LLMs become a commodity more or less and the major labs go bankrupt/get acquired by their heavy duty investors. I feel like Athropic will goto Amazon (AWS) and OpenAI may end up property of Microsoft. Google will remain Google of course so they're not going anywhere which is probaly why they won the deal with Apple.
I'm pretty confident it's Gemini behind the curtain for Siri.
They wanted to be. Thus their investment into Siri in the first place. A revolutionary system - for year 2011. As well as bleeding edge advances in computational photography, photogrammetry, etc.
They just completely failed at capturing the modern chatbot wave.
They tried to catch up multiple times and, ultimately, gave up on doing it in house. Not because they didn't try, but because they tried and found themselves lacking.
Let's be very real about Siri... it is a sad implementation of what has a lot of potential.
Talking to my HomeKit, turning on and off lights, sometimes, other times, "I don't know what lights you are talking about", "I can't find those lights" even though they're visible and reachable and controllable about the app.
"Do X" "Okay", "Do [very very similar synonym for X]" "I don't know what you're talking about."
CarPlay and Siri, unless you make sure permissions match, with CarPlay giving you navigation, press the Speech button, "Find me the nearest Starbucks." "I'm sorry, I don't know where you are".
It has nothing to do with "not being in the LLM laboratory business". I get, and agree with that. But Siri has been around for 14 years at this point and is barely more than a simple voice control for alarm clocks and timers and "play music", at this point.
I can't wait for the moment Apple realizes that hardware makers will also get eaten by AI. Who needs a fancy and expensive macbook or iphone when all you'll really need is earbuds with an internet connection to talk to the AI that's hosted wherever, which will do everything you ever want it to just by saying so. No keyboard or screen required to get a result, no real local computing hardware necessary. If the result is visual just tell it to display it on your 65" hi-res television (which Apple doesn't make). Maybe the market for earbuds is going to sustain them in the future?
People want to host their own AI and it will become good enough so most will do that instead of paying for a subscription.
Voice-only input to a cloud model with just a screen to show you what it's doing sounds like a nightmare. Why not subscribe for the TV hardware as well as the subscription, take it up a notch on the own-nothing.
You are talking about maybe 0.005% of the whole population of the earth when you say the phrase "self hosting".
My wife is part of the other 99% and she's already talking to a chat prompt for 90% of her computing needs. The fancy laptop we bought her a year ago sits collecting dust. She is Apple's target market - not the nerds that get a boner about "self-hosting".
Yea, but if I can get a ChatGPT-like experience from Siri AI for free, why would I pay OpenAI.
Now it remains to be seen if Siri AI will deliver anything close to a ChatGPT-like experience. But if they did, for the consumer segment that isn't using LLMs for agentic work and just ask it questions from time to time, I can't imagine one textarea has engendered some huge amount of brand loyalty over another.
Traffic from Siri to the web is much higher than traffic from OpenAI, generally. It's the default. People installing ChatGPT takes work. And some of that traffic is also coming from Siri today… It won't after this launches.
Because Siri defaults to dumb search much more often. While ChatGPT sucks up the search results and gives its own answer.
Which either terminates the session, or goads the user into asking a follow-up question, improving retention - the user doesn't leave the app either way.
iPhone 8 shipped an NPU in 2017 - among the first consumer electronics vendors to put dedicated AI acceleration into something. iPhone X then took advantage of that with TrueDepth and the infrastructure around it.
There was a real push from Apple at the time to enter the AI game - mainly for image processing purposes, which was the mainstream flavor of AI at the time.
For the same reason you don't run "4+6" on a calculator.
External tool call has an overhead. It requires a round trip into an external tool. It requires an LLM to run in agentic autoregression - it can't be used in prefill.
Which means that having native arithmetic capabilities is useful. Forward pass arithmetics are an LLM version of quick mental math.
An LLM can read "#define SILLY_TIME_CONST (3*20*60*60*1000)" and have "SILLY_TIME_CONST is 60 h expressed as 216000000 ms" already cached by the end of the line, before it even emits its first token.
The future needs more AI compute. No one has enough AI compute.
Memory chip vendors are betting hard on this being a temporary state of affairs that doesn't last, and doesn't warrant commissioning a shitton of new memory foundries.
Musk is betting hard on this staying that way, and is putting the next Colossus into the last place not corrupted by NIMBYs... SPACE!
Launch isn't profitable simply because ongoing Starship R&D is eating into it. A lot of opex, capex, and pre-revenue.
If they start running Starship anywhere near the way they do Falcon 9, it'll flip into profits. A lot of big bets SpaceX made ride on Starship coming online. I'm honestly surprised Starlink is already so profitable without it.
One of their big named bets includes: orbital datacenters. Which puts this specific deal into perspective.
80% of the space launch business is putting starlink satellites into orbit, so it's all internal funny money. They could very well be letting the space launch business take losses to make the satellite internet business look better (only profitable part of the whole thing).
Wasn't starship supposed to be funded by the NASA contract?
Well it would be a lot easier if those companies wanted to build them in uninhabited areas in the middle of nowhere with no infrastructure. Somehow they don't want to do that...
SpaceX has recently started pitching itself as an orbital datacenter company.
If you buy into that business model (or pretend to), it makes sense for SpaceX to start selling compute early. Their "earthside compute" clients of today are "skyside compute" clients of tomorrow.
A part of Musk's old pitch for Starlink was: space-based solar makes perfect sense for powering space assets, and no sense whatsoever for powering Earth assets. So you have to find a way to use that power in space to do something economically useful. Comms were the only scalable way to do that, so Starlink it was.
I can see how space-based datacenters would follow the same logic. If SpaceX can make them economical, that is. There's no guarantees of that - but if anyone at all can make space-based datacenters economical, it's SpaceX.
When I hear space I think "that's the perfect location for a data center", since data centers are lightweight, small, require little power, don't need human intervention, have lifetimes measured in decades and don't have to reject heat. Since space easily satisfies these requirements, space is an ideal deployment location for data centers.
Isn't the Vegas Loop just a car tunnel? As far as I know, there aren't any actual hyperloops[1] involved, just a narrow highway, even if they deceivingly brand it "Loop".
It seems off at first glance but actually appears to work out if you do the math. You can model a solar panel as a flat, opaque rectangle. You can calculate power generation and equilibrium temperature for it based on surface area. If you require additional radiative surface area to achieve the desired equilibrium temperature you can place a flat triangle orthogonal to and behind the solar panel in its shadow.
Compute is "free" at that point because waste heat is coming out of the total energy flux which was already accounted for (because we modeled it as opaque).
Of course swapping out the equipment poses a bit of a challenge. The "helping hands" rate is entirely unaffordable and wait until you see this new DC's physical access policies. 0/10 would not rack with them again.
> if anyone at all can make space-based datacenters economical, it's SpaceX
Let's hope burning ten thousand tons of toxic e-waste annually in upper atmoshphere never becomes economical. Or mankind gets to senses and bans externalizing your e-waste problem by burning in atmosphere...
> ...burning ten thousand tons of toxic e-waste annually...
Expressing water usage in gallons makes it seem really large, too. NASA says[0]:
Scientists estimate that about 48.5 tons (44 tonnes or 44,000 kilograms) of meteoritic material falls on Earth each day.
If we assume that they're all the heavier v2 units, the total mass of the orbital portion of Starlink is ten point four tons. [1] If we assumed that they lasted one year (instead of the five that they're reported to last[1]), then over the course of a year, Starlink would dump six hours worth of asteroid collisions into the atmosphere.
I think we'll be fine. Pour all that frustrated energy you have into substantially reducing the amount of incredibly hazardous d-waste [3] big commercial operators burn up into our atmosphere, instead.
I think you missed a factor of 1000 somewhere in there: Each satellite weighs about 1 ton, there are about 10,000 of them. That is 10,000 tons in orbit for the constellation, not 10. Assuming a 5 year decay, that's 10000/5/365 ~= 5 tons / day. Still about 10% of the natural incoming material, but considerably more than your "six hours worth per year".
> I think you missed a factor of 1000 somewhere in there... there are about 10,000 of them. That is 10,000 tons in orbit for the constellation, not 10.
I did. what. the. hell? Maybe my swiss-cheese brain read the "," in 10,413 as a decimal separator? I guess that's what I get for posting while old. Thanks for the correction and supporting arithmetic.
Though, I still stand by my "please for the love of everything, get to complaining about CO2 because this thing you're complaining about is a damn nothingburger" conclusion. (I am sufficiently aware to notice that that you're not OP, so the "you" in that pseudoquote is not directed at you.)
> Meteorites don't have any of toxic chemicals used to build a server with ... solar panels
Right. The toxic chemicals found in solar panels known as silicon, aluminum, copper, and trace amounts of lead. These chemicals are only found in fuming Earthbound laboratories, and are nowhere else in the universe.
You think the toxity and dangers of materials are defined by the atoms the materials are made of rather than the molecules? Swimming in a ocean of Hydrogen-Oxygen salted by Sodium Chloride must feel dangerous..
Also the toxic fumes from burning batteries don't really come from the lithium but everything else the batteries are made of.
Can anyone explain how the thermals will work? One of the biggest challenges on Earth is cooling the data center, and it's at least as challenging in space.
The earthbound equivalent would be strapping each chassis to the back of a dedicated solar panel and having the panel double as a giant heat sink. The problem is that doesn't work on the surface due to (at least) rain, the day/night cycle, and the cost of real estate.
But it doesn't matter since in this scenario each chassis is powered exclusively by the respective panel. How hot does a black panel sitting in the midday sun get? That's your equilibrium temperature. As long as it's within the operational limit of the device there's no problem.
The reason earthbound DCs are difficult to cool is because of density. When you match up panels to devices and shelter in their shadow you no longer have anywhere near the same power density.
What isn't right? I pointed out that if you adhere to the same power density then cooling is no longer a challenge on earth (in reply to the observation that cooling a DC on earth is one of the biggest challenges).
For the record the equilibrium temperature in earth orbit is above freezing but below room temperature. Cooling won't be a problem at all unless you bring along a self contained power source. Heat distribution however might be - you will need an efficient yet lightweight construction to spread the heat generated in the chassis across the entire solar panel.
The idea that passive radiators don't work on the ground isn't right. I wrote it badly, the core of your argument is really fine.
The reason we don't use them is because the other options are cheaper. But passive radiators on the ground are orders of magnitude cheaper than on space because they can use convection and conduction.
> The reason we don't use them is because the other options are cheaper.
Yes, that is literally what I have been saying from the beginning. Are you sure you didn't misread my original comment?
> passive radiators on the ground are orders of magnitude cheaper than on space because they can use convection and conduction.
That statement is technically correct when comparing designs to radiate an equivalent amount of heat in the two environments.
However in context (ie solar powered computing in outer space in the vicinity of the earth) it is entirely missing the point that the problem is not a lack of surface area but rather efficient heat distribution across the already existing surface area. I have no idea how much that costs in materials and workmanship but when you're boosting things into orbit I don't think the material cost of a rudimentary heat spreader is likely to be of much concern. The weight certainly will be, but you can also get away with some incredibly flimsy designs when operating in zero g.
It won’t. It’s not supposed to work, it’s a mirage to raise dumb money. It’s way, way more challenging to cool something a vacuum. The only option is radiative cooling, which is far from being performant. The idea is as realistic as Musk previous grifts such as his digging company and there hyperloop, both absurd and supposed to revolutionize transport, both created as grifting devices and ensure public transport doesn’t develop in the US
Musk is a snake oil salesman (that’s been clear since the self-driving car promises) but he also has made a lot of people a lot of money and that’s all anybody really cares about.
None of his companies have a traditionally reasonable valuation. Is there any reason to think that’s going to change soon?
... what? Data centers are literally the original form of computer facility. How are they different from the computer rooms mainframes, etc were housed in?
There are asteroids with concentrations of precious metals more valuable than earth's entire economy. Why don't we just send up spaceships to mine them and send the haul back to earth? What country would say no to free money?
After all, it's just an engineering challenge, not impossible.
The numbers on that are at least somewhat questionable. Even ignoring that you'd crash the market (thus it's not actually worth what it first appears to be) what is the total fuel cost to adjust the orbit of the target asteroid to land the entire thing back on the earth? Because that's what you're doing bit by bit as you shuttle loads of ore back.
Now if you have space based manufacturing or fuel production on the other hand ...
That's the point. Basic rule of thumb: anytime someone is arguing that the military will fund something, they're wrong.
Its not a real argument it's just used because to most people the military is a big mysterious thing they don't understand which they think has an infinite budget for things.
There are no NIMBYs in space. No government permitting on land use. And solar power is plentiful. It's like having a dollar store Dyson sphere.
Making use of that is predicated entirely on being able to put a lot of hardware into space cheaply. SpaceX is the undisputed best at that, no one comes close. The question is whether their "best" is good enough to make space datacenters economical.
There are many Not In My Orbit people on this very page. Many current national politicians would be happy to vote AI out of orbit today. Space is not an escape from earthly politics.
I am surprised how many people say that there is no reason to put data centers in orbit, when, at the same time, data centers are becoming the hated thing du jour all over the US and politicians left and right (but mostly left-of-center) are touting bans and restrictions to their electorates.
It is definitely to escape most political pressures on Earth. They will never be able to sidestep the US feds, but aside from an open war with China or Russia, all other authorities are out of the game when it comes to space.
People don't want to live near data centers. But companies find it logistically cheaper and easier to keep proposing to build them near existing towns and infrastructure, and then deal with regulatory fights rather then picking an isolated area and running an extension of high voltage lines out to them.
Which tells you something about why space data centers makes no sense.
"rather then picking an isolated area and running an extension of high voltage lines out to them."
Does not this usually mean extending/upgrading roads and other infrastructure as well?
IDK how this works in the US, but in most of Europe, a "linear" project like this, which crosses multiple jurisdictions, usually runs into more resistance, not less. The multitude of people and special interests along the line compounds.
In some places, special legislation has been enacted that exempts such linear projects from detailed review and opposition, otherwise pipelines, grid upgrades etc. would stall for decades in courts.
The data link between earth and space has so much bandwidth.
There are sensors in space that send data to earth it gets processed and then the data is sent back to space then to the end user back on earth. If you do the compute in space you save the space-earth transfer time twice. Latency and availability of bandwidth are both factors.
There may be limited utility for this outside of military.
Evaporative cooling is the way it happens down on earth - and that shuttles h2o molecules from dense useful clumps like aquifers and rivers to a less useful form spread out in the air. But evaporating h2o isn’t an option in space afaik - since there’s a shortage of air to take up the h2o. In fact I think radiative cooling is the only actual option in space.
Radiative cooling. When something gets hot, it begins to radiate that heat away - black body radiation.
Fucking hell - do you all think ISS is cooled by hopes and prayers?
Starlink V3 sats have to dump ~20kW of pure waste heat just to exist. Going from that to the stated 100kW is an engineering task, not some sort of impossible arcane rite.
That's the neat thing: you don't, or at least not in the megawatt range. A kilowatt can be done with radiative cooling but doesn't get you far with a hypothetical datacenter satellite.
No; if you try to do this you don't launch in the first place because the amount of servers required to be useful can't be cooled within your payload budget.
My job is mostly worrying about cooling paths, maintenance, power, heat transfer, lifetime of GPUs, and high performance networks. NVIDIA partner. I can drive to the datacenter. This stuff BARELY works here on Earth. Especially thermal issues.
Looking forward to watching spacex defeat physics.
Starlink V3 bus already has to dispose of ~20kW of waste heat from the electronics - because RF amps aren't that great at what they do. That's a ~2020 server rack, in SPACE.
Going from that to a 2026 server rack is engineering, not magic.
Not to mention that those capabilities are inherently dual use. If you know how to write C safely, you know how to spot unsafe C.
reply