Forgot your password?
typodupeerror

Comment Re:There is one reason to do this, and it isn't co (Score 1) 113

The advantage of space is perhaps (for now) less regulation and immediate power availability. It seems it's never going to be cheaper since the lifetime cost of a datacenter GPU is primarily the GPU itself, not the lifetime cost of electricity or cooling (both free in space) which only add a total of about 20% to that.

I'm not sure about being impossible to turn off - lawmakers could always insist on that, and an adversary like Russia could always kill the satellites, or jam communications. I guess being in space would at least prevent low-tech drone attacks.

We're never going to have the launch capacity to maintain datacenters in space as an alternative to terrestial. If you do the math even SpaceX's projected 40K Starlink satellite constellation will require 2-3 launches per week just to replace routine end-of-lifetime satellites. To repace terrestial datacenters you'd probably need to 100x that number (google rack shipments to get an idea of the scale).

Comment Who is buying them ? (Score 3, Interesting) 59

Boston Dynamics has always made the most capable humanoid robots, but it's not clear what the market is for them. Their dog-like Spot seems to sell to police forces and military and serves a purpose to go where no human can safely go.

Musk is projecting a market in the millions for his TeslaBot, but various Chinese companies like Unitree already have humanoid robots for sale and don't seem to be finding much of a market for them yet outside of quirky hospitality/entertainment usage. Some companies evaluating them for industrial use, but AFAIK no large scale adopters.

Comment Re:correcting repetitive reasoning? (Score 3, Interesting) 45

Weight values control what tokens get predicted/emitted.

Reasoning (to be understood in context as LLM regurgative reasoning) is ended by emitting an "end-think" token, otherwise it will continue.

The overly dramatic headline might have fooled you into believing that Astra was programming Qwen directly, as a possible alternative to training, but if you actually read the article the reality is just that Astra fixed a bug in Qwen so that it predicted (emitted) the "end-think" token at the appropriate time.

Comment Re:Not as impressive as it sounds ... (Score 1) 45

The defect was noted on Qwen's model card - there is no indication that Astra figured out for itself. What Astra did was apply it's mech interp knowledge to locate and fix the source(s) of the bug.

Saying that Astra boosted Qwen's math score is like saying that I boosted the top speed of your car when all I really did was fix your flat tire (which did allow you to drive faster).

Comment Not as impressive as it sounds ... (Score 1) 45

Astra didn't do anything to specifically improve Qwen's math ability, but rather fixed an unwanted behavior of the model, noted on it's model card, of it continuing reasoning for too long when it should have stopped - failure to emit a token at the right time.

A less dramatic headline would have been "Astra fixes Qwen bug by modifying its weights". This is mechanistic interpret-ability stuff that researchers already know how to do.

Still cool, but isn't AI making AI smarter the way the headline invites you to think. It's a bug fix, but in a model's weights rather than in code.

Comment Re:Heat and Power (Score 1) 112

The TPU being used is the Trillium v6e which draws 300W.

According to Tom's Hardware heat dissipation is what limits runtime to 15 min, which is presumably why they only run one TPU at a time despite having the power (1kW) to run 2 or 3 of them, since heating would be proportional to how many are running, and max runtime would presumably reduce accordingly.

https://www.tomshardware.com/t...

It seems we have entered the era of bullshit tech "demos" - Musk's empty shell of a car that can fly for 3 sec on compressed air, and now Google's "datacenter in space" that can run 1 TPU for 15 min before overheating.

Comment Not exactly a datacenter ... (Score 1) 112

> However, the power demands of the TPU mean it can only turn one on for about 20 minutes at a time

They've got 4 TPUs (Trillium v6e) on board, but can only power one at a time, and even then only for 20min (I assume this means it runs from battery, not direct from solar).

It's a bit like me calling my desktop computer a "datacenter", except my desktop computer can run continuously, not just for 20 min at a time.

Comment Re: Is any company managing to make AI anything pa (Score 1) 96

Sure, but Google are trying, not altogether successfully, to navigate this potentially dangerous (for them) transition. It's also been a while, predating LLMs, since Google search was keyword based - they've been trying to interpret the search input for a while - to treat it like a question which is how users have intuitively expected it to behave.

IMO at this point Google is giving us the worst of both worlds, old and new, rather than embracing the switch to AI-based question answering and follow-on conversation/drill down. I guess part of the problem is that in order to provide a free (ad supported) service they need to keep cost down which means using a cheaper/less capable model, as well as the fact that until fairly recently Google's best wasn't that great.

Comment Re:If AI could actually code... (Score 1) 166

Exactly - it's the kind of app where the difficultly is all in the details and security, not in getting a prototype up and running.

No doubt a non-developer "product manager" can vibe code a prototype - I don't doubt that. I highly doubt that a non-developer will be able to guide AI to developing something secure and bug free,

Slashdot Top Deals

Save energy: Drive a smaller shell.

Working...