Google's new generation of Tensor AI chips is actually two chips, one for inference and one for training.
See full article...
See full article...
Google unveils two new TPUs designed for the “agentic era”
- Thread starter JournalBot
- Start date
infernal666
Ars Praefectus
Yet another solution in search of a problem to solve, what a giant waste of money.
Upvote
10
(39
/
-29)
It's a bit sad to see we're spending more on educating AI models than we are on educating humans.
Upvote
102
(115
/
-13)
2x the performance in 1 year (FP8 EFlops/pod size) in the same power envelope. That's incredible.
Upvote
44
(46
/
-2)
Makes sense, why would billionaires want to waste money training assets they aren't even allowed to own??
Upvote
80
(85
/
-5)
The back and forth between generalization and specialization never ends. Specialization is liked to maturity of the application, maybe we start to see some of it for AI.
Upvote
-3
(1
/
-4)
Lest we forget all the naysayers in this previous article from November: https://meincmagazine.com/ai/2025/11/...ble-capacity-every-6-months-to-meet-ai-demand.
Google is on track to double capacity.
Google is on track to double capacity.
Upvote
-4
(15
/
-19)
I know some people that put AI to good use (my brother loves Claude for coding) so I'm not saying it's an useless bubble but in my personal case I have yet to found a compelling use case for anything. Perhaps save some effort on Google searches.
Upvote
10
(15
/
-5)
The Google AI summaries arrive at the wrong answer faster than having to watch a 20 minute YouTube video to arrive at the same wrong answer so that's a definite win in my book!
Upvote
65
(70
/
-5)
When you are looking for something quite basic (for exemple how many episodes does a given show in a given season has) that short summary can be all you need instead of opening the first link or a dedicatted site.
Yes, quite low hanging fruit.
Yes, quite low hanging fruit.
Upvote
6
(7
/
-1)
Is Eflop the same thing as an exaflop (10^18 floating point ops per second)? A 4-bit floating point add/subtract/multiply takes hardly any gates, but these are still astonishing numbers. Do we know anything about what's in these chips?
PS The details of Google's arithmetic units can be found in this patent: Low-precision floating-point datapath in a computer processor. Issued in 2023. It describes how to do all the flavors of both FP8 and FP4 in one unit, with an integer accumulator, interestingly.
PS The details of Google's arithmetic units can be found in this patent: Low-precision floating-point datapath in a computer processor. Issued in 2023. It describes how to do all the flavors of both FP8 and FP4 in one unit, with an integer accumulator, interestingly.
Upvote
11
(12
/
-1)
Upvote
33
(40
/
-7)
infernal666
Ars Praefectus
I don't need an AI for that, just a mirror, you know, if I cared.
/shrugs in beard and shoulder length hair
Upvote
5
(11
/
-6)
LLM's are not AI so claiming "agentic" is just more fluff for this nothingburger.
Upvote
-5
(4
/
-9)
MilanKraft
Ars Tribunus Angusticlavius
I am saddened (but not any kind of shocked) to see the zeitgeist OpenAI is trying to manufacture with their tv ads for 20-somethings, where everyday miracles like GPT reciting a pasta recipe prior to your date (there are free pasta recipes on the internet?? unpossible¡¡¡), or GPT telling you "Don't give up, Johnny; keep doing those push-ups!" (you can insert another "JFC" here).... is resonating and becoming a thing, despite how easily every type of day-to-day information can be looked up on one's phone or laptop with a liteteral 10 second web search. I used to make fun of these ads on this very forum, but now that I've seen this, the fun is dead.
This is normally the part where I say the current crop of 20-somethings were utterly failed by their parents, also get off my lawn, etc etc.... but instead I think I'll just continue on with today's "news journey of despair."
Misanthropic is the way.
Upvote
17
(18
/
-1)
This is why I think in the long term OpenAI is probably screwed. Eventually this bubble is going to pop and then it is going to be about whoever can drive their costs the lowest, and that won't be the companies paying Nvidia billions of dollars a quarter for less specialized hardware.2x the performance in 1 year (FP8 EFlops/pod size) in the same power envelope. That's incredible.
Upvote
18
(20
/
-2)
Well, you win the comparing apples to geese competition.
In order to get a trend, you really do need more than two samples. Three begins to establish a statistical universe, but if you have a fourth that's off kilter, you blow it up. So, I'd say it's a bit premature to be calling anyone out about doubts regarding AI capacity.
Especially when AI has yet to make a fucking profit. This is just Google burning revenue HOPING people will pay enough for them to profit from. But in all the time AI has been around, that's NEVER happened.
Reality sucks that way - for them. personally, I'm fine with it. Yeah, it's going to tank the global economy with all the trillions of dollars thrown into the furnace turn out to be gone forever. But that future is a fuck-ton more likely to happen than AI becoming profitable the way they do it now.
Upvote
1
(7
/
-6)
yeah totally, i haven't seen even one use case where ai is useful.
this broke mindset leads no where.
Upvote
4
(10
/
-6)
Until some LLM provider actually backs up their product by taking responsibility for the product output instead of foisting responsibility on the user, I do not see the mass market appeal. Companies using customer facing LLMs will have to take on the liability for LLM failures. B2B works as companies will take on liability as a LLM customer. You still have the end market problem that LLM's will be commodities. Same sources of data, same methodology - they will, as now, shift around on the performance 'leaderboard', but effectively be equivalent for most uses. As commodities their margins will be poor. Apple for example spends way less than others on LLMs and simply bought access to Gemini at a tiny fraction of what Google spends ($1B vs. $185B). They could do that because they shopped around and got the best deal and Gemini was good enough. Like the internet bubble, nobody said the internet was going away. The bubble was that investment was so overhyped that many investors were burned. LLM hype is so large it will burn a lot of ordinary folks when the stock market takes the hit. LLMs will still be around though.
Upvote
21
(22
/
-1)
The individual pod’s performance scaling 10x is nearly as important as the effciency gains. Model training runs into non-linear constraints scaling beyond the rack scale interconnects within a given “pod” (Nvidia and AMD use different terms.) In no small part this is due to copper’s inherent data losses over a given length, and hence the hybrid optical/copper solutions that have become in vogue.2x the performance in 1 year (FP8 EFlops/pod size) in the same power envelope. That's incredible.
More than Moore and Semianalysis have strong deep dives on this topic for anyone interested.
Upvote
2
(3
/
-1)
I think we're all talking about whether is does useful work, not about investment valuations. If it does useful work that's all that matters in the long run, which companies win or lose is irrelevant.
There's a huge amount of defensive people who are religiously anti-AI and a huge amount of pro people who use AI all the time - for context there's ~900 million weekly chatgpt users.
LLMs are here to stay, the open models are incredibly good. For example, I use Gemma (Google) for millions of OCR tasks daily and it is incredible at scale.
Upvote
-9
(4
/
-13)
Right now one of the only use cases pointed out for AI use in the thread is to remind you to cut your hair.yeah totally, i haven't seen even one use case where ai is useful.
this broke mindset leads no where.
Not exactly inspiring for the future of humankind or that there will be some explosion in jobs because of this.
Upvote
6
(9
/
-3)
I don't know why people look down upon personal and home tasks. It's the single largest mass market opportunity in the nexus of AI and robotics - beyond replacing expensive coders, where do we impact people's lives for the better. Where we free people from domestic labor and the mental labor that largely falls upon women in the US.
Juggling tasks and appointments: school, sports, doctors, dentists, grooming, playdates, shopping, cleaning, maintenance, transportation, takes up an enormous amount of time.
Upvote
-7
(6
/
-13)
You haven't explained how the AI is a better technology for this purpose than a calendar
Upvote
29
(30
/
-1)
I'm glad to know you think so little of women that you think they can't even decide for themselves when they should get a haircut.It's an example of what a high-end personal assistant does. You match up the social event calendar with the season. Many women don't cut their hair as frequently as men do. Other examples include scheduling mani/pedi, blow outs, etc. The women's grooming market is ~$70 billion per year.
I'll say it again. Jesus fucking Christ.
E: if you felt something just now, it was my wife face palming when I told her about this.
Last edited:
Upvote
23
(26
/
-3)
You know, if LLMs could just accurately copy and paste the ingredients list from the bottom of the god damn online recipes to save me having to scroll for 5 minutes past the
Of course with my experience with LLMs, it would hallucinate some god awful amalgamation of multiple recipes with no regard for scale and you'd end up with something truly inedible.
Upvote
12
(13
/
-1)
Almost all of whom aren't paying for it and ChatGPT, as a company, has yet to be revenue positive--in spite of tens of billions of capex pinky-promise never mind all their past R&D spending.
When something is free people will use it. See YouTube. But businesses, particularly viable businesses, and charities are different things. Currently all LLM companies are VC-backed charities trying to win the monopoly before converting to actual businesses. They're all hoping to corner the market and win, and force everyone to then start paying enough to make the product actually financially viable. See also YouTube.
Which is just a bubble. And of course, cannot happen.Google is too big and diverse to give up, same for Microsoft, there are probably China-centric omni-corps that are similar. All the speculation is based on the premise of consolidation and monopoly-endpoint.
Upvote
9
(10
/
-1)
It would be interesting to know exactly how you program these chips. What is the API, how does the Linux driver look and all other similar questions.
Upvote
1
(2
/
-1)
OrangeCream
Ars Legatus Legionis
The same was true of dot.com in 1999
Yes lots of wasted effort and ruined companies, but also gems like Amazon and Google powered through the crash
Upvote
-7
(0
/
-7)
It takes a lot of time to actually DO those things - not to plan and schedule them. If people are having a lot of issues handling the average scheduling of everyday life then we need to look back at what is being taught in school. Those people need to learn better task/schedule management a lot more than they need software to tell them what to do and when.
Upvote
3
(5
/
-2)
OrangeCream
Ars Legatus Legionis
Something or someone has to create the calendar first. In case it’s not clear, AI is a tool capable of creating a calendar.
Upvote
-11
(1
/
-12)
Upvote
14
(14
/
0)
OrangeCream
Ars Legatus Legionis
Scheduling work is a huge undertaking. Doing it successfully is the difference between a good and a great project manager.
Upvote
-15
(1
/
-16)
When did we talk about project management?
I responded to a comment about the average everyday scheduling of personal chores. The stuff you need to be able to do to live on your own.
The discussion had nothing to do with professionals performing scheduling as a job.
Upvote
16
(16
/
0)