Google says its more efficient Gemini 3.5 Flash is the key to your agentic AI future.
See full article...
See full article...
Google announces agent-optimized Gemini 3.5.Flash and a do-anything model called Omni
- Thread starter JournalBot
- Start date
Speaking as a non-engineer but not an entirely uninformed idiot, AI scared me.
Agentic AI scares me more.
Agentic AI scares me more.
Upvote
42
(53
/
-11)
Antigravity 2.0 is giving me "failed to make code backend request" on install. Anyone actually get it to run?
Edit:
View: https://www.reddit.com/r/google_antigravity/comments/1thuut3/new_update_issue/
Looks like the roll out is screwed up.
Edit:
View: https://www.reddit.com/r/google_antigravity/comments/1thuut3/new_update_issue/
Looks like the roll out is screwed up.
Last edited:
Upvote
1
(3
/
-2)
It's not Star Trek they are working towards, it's The Hunger Games.
Upvote
107
(121
/
-14)
Upvote
-2
(2
/
-4)
Speed is not my priority. Accuracy and reliability is.
I will not be using any system with Agentic AI enabled. This might mean I eventually have to do without computers.
I will not be using any system with Agentic AI enabled. This might mean I eventually have to do without computers.
Upvote
77
(83
/
-6)
We live in a consumer driven economy. There's no way they would let us basically be labourless. It just doesn't work.
"AI" can never reach sci-fi levels with current technology. It just cannot. It took a breakthrough of material science to deliver us PCs that could fit on a desk. That same breakthrough needs to happen to AI, or else it'll forever be a massive waste. And with current tech, it'll never happen. Inference on the scale of LLMs is just too costly to implement.
Don't delude yourself into thinking we're gonna have Star Trek levels of AI or society anytime soon. I would almost guarantee you we'll destroy our environment WELL before getting 20% of the way there.
Upvote
58
(70
/
-12)
The people to mind aren't the economically-irrelevant minions...it is people like Peter "I do not think democracy and freedom are compatible" Thiel
Upvote
87
(87
/
0)
From the article:
If the costs of tokens is the problem, how is churning through them more quickly the solution? I didn't see anything in the article saying that 3.5 Flash is more efficient or costs less, just that it is faster. The way I see it, this new model release seems to be doubling down on, "we lose money on every token, but make it up in volume".
Maybe I missed something in my read of the article, or there is extra context here, but the problem, "generative AI is currently a money pit", and the solution up, "the new model can output nearly 300 tokens per second", don't seem to be actually connected.
If the costs of tokens is the problem, how is churning through them more quickly the solution? I didn't see anything in the article saying that 3.5 Flash is more efficient or costs less, just that it is faster. The way I see it, this new model release seems to be doubling down on, "we lose money on every token, but make it up in volume".
Upvote
54
(64
/
-10)
Who is "they"?
Do you live in a democracy, or are you living in a pure authoritarian state without agency?
Upvote
-8
(18
/
-26)
I'm not in software development so I'll leave that for the programmers to say how good AI is for the task. But for a lot of people who are trying to sell us on these AI tools, I always read into their quotes that they are useless space-fillers and that AI is basically doing their job for them. Perhaps I'm too cynical on the corporate world, where the people who love AI the most are the ones who seem to generate the most fluff-filled, content-devoid output.
Upvote
31
(33
/
-2)
Well, no and yes. Yes, we have a consumer driven economy...BUT...who are the consumers? You know how stock ownership is severely top-weighted to the point where 90+% of all trades are billionaires?
The same exact thing has happened to consumer spending in most "western" countries. Particularly the USA. The asset-owning top-10% in the USA performs about 50% of all consumer spending now. Which is why so many economic KPIs like GDP and CPI and so on--are increasingly decoupled from Normal Person lived experience. A trend foreseen 20 years ago by the marxist communists at Citibank--and their recommendation to investors and clients was to abandon working-people and pivot to only servicing high-net-worth individuals.
https://www.sourcewatch.org/images/8/86/CITIGROUP-OCTOBER-16-2005-PLUTONOMY-MEMO.pdf
That memo...they tried to retcon from the internet...because it is a pretty terrible look to say the least.
Upvote
45
(46
/
-1)
The unsupported but plausible implication is the that it means they can produce 300 tokens per second on hardware that used to produce 75 tokens/second (or whatever). So more throughput on the same hardware. To what extent that holds up depends on a lot of details that google isn't sharing.
If Google can run models that compete in performance with competitors and their old "pro" models, but cost google similar to their old flash models, that could definitely change the profitability equation.
Upvote
55
(55
/
0)
It's plausible to infer that individual sessions aren't being given significantly more processing power. As such, more tokens in the same time implies that Google has found shortcuts that let them process a given token more quickly and therefore is taking less computer (and thus less money) to generate.
Also, the issue with AI is two-fold: queries are getting cheaper to process. However, accuracy still isn't great, and the easiest way to up accuracy is by using more compute per query. So there's a tug of war between wringing efficiency from each token calculated, and using more tokens in each query response.
How much this tech succeeds ultimately depends on if they can get accuracy high enough to satisfy end users while finding enough efficiency gains to eek out a profit.
Last edited:
Upvote
28
(28
/
0)
Eh, I doubt it.
Oh, you mean financial sense for the companies to pursue. Maybe, sure.
Upvote
4
(11
/
-7)
Putting aside how terrible Google has been at shoving AI everywhere it's not needed or wanted, with opt outs buried like the best pirate treasure, and how Gemini is a pretty bad, try-hard model that will completely rebuild a website when you simply ask for a font color change, things like this are why I'm much less worried about a lot of the scary environmental impacts of AI.
Electricity and water usage fears aren't unfounded, but they were always in that realm of engineering problems that we've proven incredibly adapt at overcoming - it felt inevitable that the models would become more efficient, and that the hardware would as well, eventually significantly reducing the amount of electricity and water required, and finally trickling down to reduce the ridiculous CapEx numbers AI companies were throwing out there for data centers.
AI still has a hosts of problems - impacts on employment, AI slop masquerading as legitimate news, deep fake nudes, a huge stock market bubble - but at least solutions for the environmental issues are coming soon - maybe not as soon as we'd like, but better late than never
Electricity and water usage fears aren't unfounded, but they were always in that realm of engineering problems that we've proven incredibly adapt at overcoming - it felt inevitable that the models would become more efficient, and that the hardware would as well, eventually significantly reducing the amount of electricity and water required, and finally trickling down to reduce the ridiculous CapEx numbers AI companies were throwing out there for data centers.
AI still has a hosts of problems - impacts on employment, AI slop masquerading as legitimate news, deep fake nudes, a huge stock market bubble - but at least solutions for the environmental issues are coming soon - maybe not as soon as we'd like, but better late than never
Upvote
-14
(6
/
-20)
That was a pretty impressive display by Google.
I still don't think there's much there that will make me open my wallet yet but they gave a very strong feeling of everything continuing to speed up.
Can't help but feel retailers will hate the universal shipping cart though.
I still don't think there's much there that will make me open my wallet yet but they gave a very strong feeling of everything continuing to speed up.
Can't help but feel retailers will hate the universal shipping cart though.
Upvote
3
(5
/
-2)
Yea, use Claude Code or Cursor. Antigravity sucks, it's a try-hard that will completely re-do a kitchen when you ask it to fix a stuck drawer. Antigravity f*cked up more simple tasks than I care to count, it tries way to hard to 'guess' what else you might want (refactor an entire webpage) even when you make simple request (change a font color)
Upvote
3
(8
/
-5)
Unless a few thousand people with pitchforks are outside [name redacted]’s mansion with pitchforks, why would they share any wealth to create a Roddenbarry imagined future?
Zero bankers went to prison after the financial crisis while thousands of Americans lost their homes. Boar’s Head killed more people last year than the Manson family during their entire murder spree. I can list examples all day. It’s nonsensical to think the wealthy will suddenly decide to share for the good of all mankind because AI when they have been greedy for the entire history of mankind.
E.g. How many Marc Cuban libraries have you been to? How many Elon Musk opera houses? How many Sam Altman cancer wings are at any US hospitals?
Bill Gates* and Laurene Powell Jobs are statistical anomalies in our world - why do you expect them to become the norm?
*you know why this asterisk
Is here
Upvote
53
(55
/
-2)
Upvote
15
(15
/
0)
Utility usage isn't a engineering problem. They're policy problems. Policy allows datacenters in (e.g.) Arizona to use evaporative cooling. Closed loop cooling systems are a solved problem, they just cost more to build and operate. So, absent a legal requirement, data centers are going to build open systems that turn scarce local water into vapor. Require closed loop cooling and you eliminate the bulk of water usage.
Similarly, my electrical bill going up due to data centers is also a policy issue. Data centers, as large bulk buyers of electricity, are allowed to negotiate lower rates. End users are basically stuck with whatever rate they're given. So as rates are bid up (since supply is finite and constrained in the short run), households get hit harder than big users. Regulators could add surcharges to high users (and pipe the money towards rebates for households) to balance this out. I'd be stuck paying more for the same usage, but it wouldn't be as bad.
We're starting to see some of these ideas get enacted at the state and local level.
Last edited:
Upvote
33
(33
/
0)
Well, the implication I think is that "on the same hardware this model can produce 4x as many tokens in the same time", which should mean each token costs 25% as much. Which would be a significant change, if the quality was similar.
We'll have to wait to see if those things turn out to be true, though.
Upvote
9
(10
/
-1)
What implication? Was this hinted anywhere that this was the same hardware? (Genuine question, trying to parse headlines here)
I just would’ve assumed they rented out specialized providers like Cerberus that already provide high-speed models of 100+ token/sec today, like Opus-Fast and Codex-Flash that cost 5-6x more.
And if it was the same hardware, why does it now suddenly cost 5x more per token? [1]
1.
View: https://x.com/artificialanlys/status/2056795055512596817
Upvote
5
(5
/
0)
The damage is being done now. 50% of the voting age people are not real bright and the money behind AI is counting on that so they can damn the torpedoes and go full speed ahead.
The only real hope is for the house of cards to crumble and for some real guardrails to get put in place.
Upvote
13
(17
/
-4)
They can trot out all the rich executives they want and they can throw an entire dictionary's worth of enthusiastic terms around, but....I can't think of anything I'd need or want an AI agent for. I wouldn't give one access to my servers or network infrastructure, nor would I give one access to anything more personal, both for reasons of privacy and because no matter what the executives say, these models keep making random mistakes.
Upvote
18
(19
/
-1)
You're not actually thinking about solutions, either, but imagining a literal deus ex machina.
Star Trek ran on magic and the fiat of writers, and there is simply no guarantee the latter is possible or even desirable. In fact, automation has historically created more jobs, rather than fewer.
Technological progress does not and cannot go on forever; it will eventually end for anything under the Sun. When and where that progress ends for whatever metric you choose to employ varies, but it will end and often far short of where you want it.
If we want real, long lasting solutions, then it will be up to people to actually take action instead of fantasizing about magic disguised as technology saving the day.
Upvote
27
(29
/
-2)
I... I mean I just don't trust this envelope you've got there
Near as I've been able to tell, they might turn a profit in a few years fingers crossed
Upvote
15
(16
/
-1)
Upvote
1
(1
/
0)
Well then you will happy to know that, in true Google fashion, the new Antigravity is a completely different product than than the old one called Antigravity.
Upvote
11
(11
/
0)
Indeed, "We made our chaosbot cheaper!" is a solution to a problem that I do not have. It could be free; it could be $20/day paid to the user. I'd still never let it near a system I even slightly cared about--treat it as radioactive as a flash drive you found on the sidewalk.
Upvote
11
(12
/
-1)
I have never understood this mindset, this expectation of mindless clapping like a brain-dead trained seal. Do you expect a car website to favorably review every new model it drives? Do you expect the sports section of your local newspaper to bleatingly praise every new player signed and every one of the coach's in-game decisions?
Ars Technica writes about technology, but "about" means a whole lot more than reflexively showering with fanboi-ism. Is Engadget still a thing? I think you might find it more over there.
Upvote
39
(45
/
-6)
Or better yet, don't! There aren't any validated improvments in productivity in anything but an anacdotal way, its going to get more expensive and difficult to control costs on, no matter what model you use, and ultimatly makes you a worse thinker and engineer than learning to do what you want yourself.
Upvote
10
(14
/
-4)
Upvote
-8
(1
/
-9)