A big part of this announcement does seem to be _delegation_ in the background; they give the example of web search but that could be any tool. I haven't tried it yet either but sounds like they've found a reasonable UX that mixes that sort of high-and-variable latency tool calling (potentially with agentic loops) with a continuously speaking live voice.
Aside from the issue of platform owners (Apple, Google, Microsoft) offering storage sync as an integrated feature, which others have pointed out, the other reason growth is limited is that filesystem storage and sync thereof has become less critical over time. Our apps increasingly do cloud-native walled-garden storage: docs in Google Docs / Notion / Confluence, mocks in Figma, etc. And code was cloud-native with version control much earlier.
I need sync for just photos on my phone (which Apple or Google are better for), and a small number of esigned PDFs and tax documents (for which any provider's free tier suffices).
Looks like the best display you can get in laptops at this price: 2408x1506 resolution, 500 nits, antireflective coating (!). And bonus points for no silly notch.
I guess it could warn about it but the VM sandbox is the best part of Cowork. The sandbox itself is necessary to balance the power you get with generating code (that's hidden-to-user) with the security you need for non-technical users. I'd go even further and make user grant host filesystem access only to specific folders, and warn about anything with write access: can think of lots of easy-to-use UIs for this.
Pretty close to Gemini 3 Pro Image (aka Nano Banana Pro) in most benchmarks, even without thinking+search, and even exceeding it in 2 most important ones of 'Overall Preference' and 'Visual Quality'. I'm excited about the big jump in Infographics/Factuality (even without thinking+search; I'm surprised that text+image search grounding doesn't make an even bigger dent).
Agree, can't wait for updates to the diffusion model.
Could be useful for planning too, given its tendency to think big picture first. Even if it's just an additional subagent to double-check with an "off the top off your head" or "don't think, share first thought" type of question. More generally would like to see how sequencing autoregressive thinking with diffusion over multiple steps might help with better overall thinking.
Surprisingly big jump in ARC-AGI-2 from 31% to 77%, guess there's some RLHF focused on the benchmark given it was previously far behind the competition and is now ahead.
Apart from that, the usual predictable gains in coding. Still is a great sweet-spot for performance, speed and cost. Need to hack Claude Code to use their agentic logic+prompts but use Gemini models.
I wish Google also updated Flash-lite to 3.0+, would like to use that for the Explore subagent (which Claude Code uses Haiku for). These subagents seem to be Claude Code's strength over Gemini CLI, which still has them only in experimental mode and doesn't have read-only ones like Explore.
"the value of a human eyeball" / attention is and always will be the limited resource. But I wish the way the economy worked wasn't that attention is sold for money, which makes money the moat, and sets a floor on how low-priced things can get for customers too. Is this really the best the economy can do? Or is it possible to have a fair LLM-based search engine that matches customer need description with stated product descriptions from providers (while weighing customer reviews, etc)?
Hmm the whole point of checkpoints seems to be to reduce token waste by saving repeat thinking work. But wouldn't trying to pull N checkpoints into context of the N+1 task be MUCH more expensive? It's at odds with the current practice of clearing context regularly to save on input tokens. Even subagents (which I think are the real superpower that Claude Code has over Gemini CLI for now) by their nature get spawned with fresh near-empty context.
Token costs aside, arguably fresh context is also better at problem solving. When it was just me coding by hand, I didn't save all my intermediate thinking work anywhere: instead thinking afresh when a similar problem came up later helped in coming up with better solutions. I did occasionally save my thinking in design docs, but the equivalent to that is CLAUDE.md and similar human-reviewed markdown saved at explicit -umm- checkpoints.
Quite a premium for speed. Especially when Gemini 3 Pro is 1.8x the tokens/sec speed (of regular-speed Opus 4.6) at 0.45x the price [2]. Though it's worse at coding, and Gemini CLI doesn't have the agentic strength of Claude Code, yet.
Love it. Wonder if it's viable for citizen journalism in warzones and areas of civil unrest, with the larger size of photos (and short videos), given the inherently slow transfer rates and battery life implications of going thru multiple hops before Internet-exiting the area that's otherwise Internet-offline. What's the back-of-the-envelope math here on viable bandwidth?
Wifi obviously has higher bandwidth, but I guess it isn't viable as a mesh, or is there any trick with turning on/off hotspots on phones dynamically that'd make it viable? (Afaik older phones made you pick between being a hotspot or being a regular wifi client, but at least some newer ones seem to allow both simultaneously.)
I'm definitely hoping for a future with wider support for C2PA (content credentials on images) on phone cameras to make these photos power citizen journalism. So far Samsung S25 and Pixel 10 support C2PA in the camera hardware: need other phone makers (especially Apple) to get on board already... if you're an iPhone user, please help yell at Apple support etc!
Aside: I registered a domain and plan to build a citizen journalism news feed for such photos (and uncut videos). I see it as the antidote to Instagram et al's feeds that're full of AI slop (and plenty of fakery even before AI-generated imagery got big). And it's essential to truth, democracy and ultimately (maybe I'm too idealistic here) peace. Aside to the aside: wish some of us techies banded together to build "peace tech" as a new sector in tech, DM if interested in brainstorming or working together.
Sounds like antirez, simonw, et al are still advocating reviewing the code output of these agents for now. But presumably soon (within months?) the agents will be good enough such that line-by-line review will no longer be necessary, or humanly possible as we crank the agents up to 11.
But then how will we review each PR enough to have confidence in it?
How will we understand the overall codebase too after it gets much bigger?
Are there any better tools here other than just asking LLMs to summarize code, or flag risky code... any good "code reader" tools (like code editors but focused on this reading task)?
Most car manufacturers made this mistake because they started mimicking the then leader for innovation (and customer satisfaction), Tesla, too much.
General cautionary tale: just coz a company is successful, doesn't mean it's doing _everything_ right. Plenty of folks who love their Teslas would prefer a few more buttons (and door handles on the inside, etc) if given the choice. Could say similar things about some choices Apple made.
"Almost anyone can prompt an LLM to generate a thousand-line patch and submit it for code review. That’s no longer valuable. What’s valuable is contributing code that is proven to work."
I'd go further: what's valuable is code review. So review the AI agent's code yourself first, ensuring not only that it's proven to work, but also that it's good quality (across various dimensions but most importantly in maintainability in future). If you're already overwhelmed by that thousand-line patch, try to create a hundred-line patch that accomplishes the same task.
I expect code review tools to also rapidly change, as lines of code written per person dramatically increase. Any good new tools already?
Have you tried them with providing a grounding resource, e.g. attaching a file to ChatGPT or NotebookLM? Yes need some human expert to create (or curate) that grounding resource in the first place, but LLMs handle the rest well: presenting info in different ways and paces, interacting with the learner like a tutor, etc.
End of an era: video (with broadband Internet penetration) was the best tool we had for 15+ years. But LLMs are now good enough, including in image+infographic generation and factuality (especially when grounding resources are provided... which is where human experts still matter). I think video is now better only for learning physical hands-on skills... and those videos tend to be on YouTube rather than on Udemy or Coursera.
Coursera's model will still survive for a while, given people's desire for branded credentials (university degree credits or company-branded certificates)... until the university bubble bursts too in a 10+ years. Start of trend: https://www.nbcnews.com/politics/politics-news/poll-dramatic...
A bit of a plug: we tried building a consumer business, with a learning experience built atop these LLMs: https://uphop.ai/learn . Still offered for free to consumers, but we're now succeeding much better on B2B ("you either die a consumer business or live long enough to become B2B" was v true for us).
Glad to see big improvement in the SimpleQA Verified benchmark (28->69%), which is meant to measure factuality (built-in, i.e. without adding grounding resources). That's one benchmark where all models seemed to have low scores until recently. Can't wait to see a model go over 90%... then will be years till the competition is over number of 9s in such a factuality benchmark, but that'd be glorious.
Big knowledge cutoff jump from Sep 2024 to Aug 2025. How'd they pull that off for a small point release, which presumably hasn't done a fresh pre-training over the web?
Did they figure out how to do more incremental knowledge updates somehow? If yes that'd be a huge change to these releases going forward. I'd appreciate the freshness that comes with that (without having to rely on web search as a RAG tool, which isn't as deeply intelligent, as is game-able by SEO).
With Gemini 3, my only disappointment was 0 change in knowledge cutoff relative to 2.5's (Jan 2025).
It's all about the chip economics. I don't know how the _manufacturing cost_ of Google's TPUs compares to Nvidia's GPUs, for inference of equivalent token throughput.
But at the moment Nvidia's 75-80% gross margin is slowly killing its customers like OpenAI. Eventually Nvidia will drop its margins, because non-0 profit from OpenAI is better than the 0 it'll be if OpenAI doesn't survive. Will be interesting to see if, say, 1/3 the chip cost would make OpenAI gross margin profitable... numbers bandied in this thread of $20B revenue with $115B cost imply they need 1/6 the chip cost, but I doubt those numbers are right (hard to get accurate $ numbers for a private company for the benefit of us arm-chair commenters).
"complementing the Neural Accelerators in the CPU and GPU" seems to be a misprint; I don't believe they have the accelerators in the CPU too.
Still super interesting architecture with accelerators in each GPU core _and_ a dedicated neural engine. Any links to software documentation for how to leverage both together, or when to leverage one vs the other?
Former Googler, though started my SW career by learning on HN!
https://in.linkedin.com/in/zohairhyder/
Email: same user id as HN @ my former company's consumer email domain