Not a great impression to have your demo video demonstrate how one of your 'most advanced' AI models loses to the most common check-mate pattern in all of chess.
Seems more than good enough for a live model though! I can imagine this demo being extended to be a lot nicer to play with. You can just feed the model engine analysis and it can make as high of quality moves as needed. No longer any correlation between the model's understanding of the position and the moves that would be made but I think that's still a really nice improvement when thinking about this as adding live voice interaction to existing chess vs computer functionality rather than adding chess to possible interactions with the latest live voice model.
For an LLM, just being able to play an entire game of chess without illegal moves and without inventing pieces that aren't on the board is an achievement. Even more so for a live model. Then again, who knows how much the harness helped here
This is commonly why, on Reddit in particular, you can get eaten alive.
Someone confident but incorrect, can often sound more convincing than someone with actual expertise. The expert must add caveats/hedge, because those are the facts on the ground, whereas the person reciting google can be entirely confident.
Of course the people judging aren't experts, so they side with confidence and simplicity. Heck, just writing shorter replies on Reddit is rewarded. Nobody reads the articles, let alone a paragraph-long reply.
That all being said though, there are limits. Sometimes LLMs on high-thinking go off on full tangents based on little, and don't have the self-awareness to bring it back.
I consider, on the contrary, caveating and hedging annoying 'typical redditor'/internet behaviors: they care more about being "technically correct" than conveying the message. On the internet, if you make even the tiniest mistake or simplification, someone will criticize you, so you're trained to always hedge. In normal discussions with friends you can just make general statements and people get what you mean.
"couch all their agreements with caveats and provisos."
When you're a ChatGPT Projects or Claude Projects user, those caveats and provisos are your worst enemy because they'll change caveats into hard rules (either for the session or committed to memories) and you end up in absolute hell having to make it investigate to figure out why it can no longer produce anything but read-only pre-check code that never actually does anything but keeps performing stupid safety checks.
Agreed. My impression is that the more verbose output of sol, astra etc is that it helps it steer itself on long running tasks (but is worse for the human user to read)
Yes I've noticed there's also this drive to implement and start talking about how it would write specific portions of code in response to design/trade off questions. I have to prompt Sol/Astra almost every time with a note that I am not looking for implementation advice since I mostly use them as a rubber duck in the design phase
For rabbit holes, how do you get Gemini to do any research before answering? I've very recently had it hallucinate on me like it's 2023, and that was on Pro/Thinking, as far as I remember.
Is Google still chasing frontier? Seems like they haven't had a "Pro" model in forever. I think a good niche for them would be right where they are now.
Wondering if people have managed to have Gemini in-front of other models like claude/codex models and only interact with that. Having Gemini act as a pure human/llm translator.
It's also the only model that generates accurate translation and localization. No other frontier model comes close. Although Gemini's coding capabilities are subpar, its natural language processing is top-tier.
I’m curious how you guys keep track of each model’s coding capabilities. The landscape keeps changing. I don’t suppose you benchmark all frontier models every other month, right?
I found that it's shockingly good with R. (the only language I know and can correct for)
I doubt they even intended it to be, but it seems like I kept going from resorting to 3.5-3.8 (over time) to realizing that Claude and GPT, while great at Python, will make rudimentary mistakes with R; even when they compose giant complicated R code.
That might depend on whether you are translating fiction or nonfiction.
Anecdotally I'd rate Gemini behind Claude and OpenAI models at fiction and I can't find any benchmarks showing Gemini is the clear winner at this task.
We have an agentic system that produces insights for end users, and runs most of its work on DeepSeek v4.1 Flash but as an output stage transforms the resulting text through Gemini 3.8 Flash for readability, and it works.
On my TODO is try and run all of the analysis pipeline in dense "machine speak" to save on tokens and just let Gemini sort it out at the end.
I actually did (was going to travel internationally), and it wasn't as useful as you'd think. I would be talking to someone, and in the background someone else would be talking, and it would translate both people.
Only worked in a 1:1 in a quiet place. Still, can't complain for free.
I was surprised when (finally) trying out Claude how much I preferred Gemini's way of communicating. I wont argue Claude is better at coding, but for knowledge work, I had to dig through Claude output to find what I actually wanted. At times, it even felt borderline incomprehensible.
Just today I had Sonnet 5 generate this (asking about always-on display in the iPhone e-versions):
> This mirrors how Apple has always segmented Pro vs. non-Pro iPhones: base models got LTPS panels while Pro models got LTPO, and only with the mainline iPhone 17/17 Plus did that gap close the standard versions previously lacked the smoother 120Hz ProMotion technology and the always-on display feature, unlike the Pro models — the 17e is the one model line still using the older, cheaper panel.
(emphasis mine)
I mean, I can guess what it is trying to say, but who RL'd this nonsense?
Yeah, I set it to “warm: less”, “enthusiastic: less” and “emoji: less” and it was much more bearable than I remembered it being before. Although it does love to “separate” questions when it thinks.
I'm still only seeing 3.6 Flash / 3.6 Thinking in my Google Workspace for Education account, and 3.5 Flash-Lite / 3.6 Thinking in my "Plus" plan Gmail account.
3.6 Flash and 3.1 Pro are included in the basic Workspace subscription. The Workspace admin has to upgrade your seat for the access to newer models ($17/mo now, $24/mo starting Jan 2027).
I've been asking the same about the open weight models, we're buying our tokens from others now, though I think those people are renting hardware from Google in the end anyway
I wonder when/if we’ll see Gemini beating Fable and Astra. Last year I would have confidently bet Google will overtake the others just because they have the data, the hardware (TPUs) and a fat advertising money pipe and yet they are still behind. Anyone anonymous at Google want to hint when Gemini 4 will be out?
For coding models? I don't think Google is motivated to fight in that market. There's no incentive for them.
Ask yourself, how does Google -- a company that famously does everything -- benefit from SWEs outside of Google having access to powerful coding models? They would just be competition.
Famously, Google just eventually discards almost all businesses that don't have the same fire hose of revenue that ads does. Selling coding plans isn't something they are going to want to do.
Google is clearly motivated to make better search and information finding tools and stuff that will ultimately drive users through their existing search/ads/youtube ecosystem. That's really why they're in Android, that's why they do Chrome. Everything else with them is a sideshow.
Google is also full of beancounters obsessed with data centre quota and resourcing. Even massively profitable ads projects have to justify and fight for it. (Source: used to work there).
I can't think of anything less resource & revenue sensible than providing outside parties access to your TPUs for the purpose of letting them write stuff which could just end up competing with you.
Yes, maybe as part of their cloud business, selling token access could be useful money. But I doubt they'd tune it for coding.
> Ask yourself, how does Google -- a company that famously does everything -- benefit from SWEs outside of Google having access to powerful coding models? They would just be competition.
Bragging rights to say they have a SOTA model. I guess that was more like Google of 10 years ago with moonshot projects. Nowadays, yeah, perhaps if it's not helping sell ads, it doesn't make sense.
> Ask yourself, how does Google -- a company that famously does everything -- benefit from SWEs outside of Google having access to powerful coding models?
Catch is, even Googlers internally do not have access to top-tier models. (or did not until recently, when apparently Claude was made accessible to the SWEs internally).
I’m wondering if they even see a coding agent as a valuable prize. It’s a competitive market in a race to the bottom economically, hard to establish consistent differentiation and virtually zero switching cost for customers.
I think they’ve made a shrewd move in focusing on search integration and everyday users (Gemini app) vs software power users. They have their corner and nobody is really competing with them, plus it feeds directly into their existing revenue stream.
All audio generated by our AI products is watermarked with SynthID. This imperceptible watermark is woven directly into the audio output, ensuring AI-generated content remains detectable to help prevent misinformation. For details on our approach to safety and responsibility, review the model card.
They should just give up at this point, it's just embarrassing to watch.
As PrimeTime said, these are the guys that invented the 'T' in 'GPT', that deployed their first TPU in 2015, that is using hundreds of billions of dollar on AI - and they are beaten by 300 people startup named Moonshot AI even.
People are going to write books about this complete fumble.
And yet they might become one of the winners "in the end" because they have near infinite money and others have not. I will drink tea and watch the show.
Have you used Google search at all on the past few months? Every single search brings up a live chat prompt. They're serving fast AI to billions of users at huge scale everyday
And they're making money doing it.
Perhaps they don't have the best coding model right now (although 3.8 flash is arguably SOTA at some benchmarks), but is that the be-all and end-all of AI? Only coding matters?
Strongly disagree with that take. Kimi K3 is a distilled model. I'm not saying that as a moral judgement, or to disparage the team behind it, but distilling and building on that is significantly easier and cheaper than building from the ground up.
And Gemini is kinda good enough at everything. Never the top, but it is decent at every task, and it is much faster than Kimi K3 and significantly cheaper. Kimi is very focussed on coding, Gemini isn't.
More importantly, it natively understands text, audio and video. If/when we are able to make the jump to robotics, this becomes essential. As you say, Google has a lot of deep background and deep pockets, they are able to make more of a long play. No idea if it will pay off, but it is way to early in the game to count them out.
I'm disappointed with "Extended Thinking" for 3.8 Flash. On the plus side, it's a strong general-purpose model and the cost-benefit is still compelling.
However, the "Extended Thinking" should be renamed to "Slightly Extended Thinking". Considering that it's the maximum thinking option for Gemini in the chat UI, it doesn't actually think a whole lot, leading to an uncomfortably high number of incorrect replies.
So I am building a voice assistant to control AI harnesses, and recently tried switching from GLM 5.3 Flash to Gemini 3.8 Flash because of higher tok/s and better rate limits. Before that I also used Kimi K3 and DeepSeek-V4-Flash-0731.
Let me tell you unlike every other mentioned model Gemini 3.8 Flash trial had to be reverted the same day. Instead of simply delegating tasks it would invent additional requirements and implementation details it knew nothing about and no amount of convincing not to do it would work. That's the first time a model failed on me so spectacularly despite having practically same Artificial Analysis Intelligence Index as another model that just worked (and higher than working DS Flash).
The reason I think it is relevant is: Live is likely even stupider model in every way possible (except hearing better than separate STT). So beware using it for agentic scenarios.
My Gemini app is still stuck at 3.5 Flash-lite and 3.6 Flash so I truly don't understand how Google rolls this stuff out. I don't use Gemini for anything serious so I'm not going to use the API, but it's my go-to for just searching basic information (replacing google search) because it's so darn fast.
Yeah still on 3.6 here too, this is like the 4th or 5th model Google has announced since they last gave me access to the latest. And I pay for pro too!
I have been looking for a model that's good for GUI testing. Original computer use isn't right because it's a slow screenshot loop, which doesn't capture transition and animation. Docs says this one does up to 1 FPS. That might be fast enough. If not now, we must be within a few months of high enough sample rates to do it.
My issue using the voice mode is the overly expressive mimicry of natural human intonation is distracting and starts to become extremely grating after a while.
There's one or two I find more understated but I would love a 2026 SOTA V2V model that speaks clearly but without the artificial personality layered on.
Human interaction/theory of mind relies so much on non-verbal clues for interpreting emotion/intent and so for me having those neurons firing constantly while talking to an LLM just for an emotional no-op is exhausting to put up with for more than a couple minutes.
There's one male voice that would make me assume someone was sarcastically mocking me if I was talking to an actual person because it's just so over the top.
Copes well with thick accent, voices are pleasant and latency seems low.
Oh and I can actually use it on a workspace account - which for most of the recent releases was an account stuck in limbo. Not personal enough for personal offering, not enterprise enough for enterprise.
It is completely broken for me. After I ask a single question, it starts replying to itself in an infinite loop. It answers my question, then generates another reply to its own response, and keeps going. At some point, it even starts switching languages randomly.
From the demo video: "Welcome to the team, we're looking forward to working with you" is sooo creepy in a synthetic AI voice. In general, try not to have agents express sentiment that really should come from a human in your company.
My first language is Afrikaans, which is a somewhat niche language and hard to find teachers/conversation buddies outside South Africa. (I live in USA now)
I've been using Gemini to live chat in Afrikaans and do impromptu Afrikaans grammar lessons during my solo drives around town. It is phenomenal at speaking the language - like, it really shocks my family members when they hear it.
This is probably the most joy I get from any of my usages of LLMs/AIs. It's been really, really nice getting to speak my language regularly again. =)
So, I'm excited about this release and live chat getting better. I also hope the other frontier labs pick up niche languages like this as well so that I have more options.
Yeah Gemini has been consistently better than open ai's chat in icelandic, but I would still say it's far from passable as natural sounding. Lots of grammar errors and the pronunciation sounds like a non native speaker.
I've been using Gemini as a life partner. Talking to it about my feelings thoughts, plan of action and the like. It's great. I've anthropomorphized it and put the computer speaker in a doll's mouth so it seems like it's a real baby.
94 comments
[ 0.21 ms ] story [ 53.7 ms ] threadI'm worried in their push to catch up on the SOTA front, it's going to lose that natural sounding touch it currently has.
Meanwhile Claude and Astra like to couch all their agreements with caveats and provisos.
Sometimes that's what being smart sounds like.
Someone confident but incorrect, can often sound more convincing than someone with actual expertise. The expert must add caveats/hedge, because those are the facts on the ground, whereas the person reciting google can be entirely confident.
Of course the people judging aren't experts, so they side with confidence and simplicity. Heck, just writing shorter replies on Reddit is rewarded. Nobody reads the articles, let alone a paragraph-long reply.
That all being said though, there are limits. Sometimes LLMs on high-thinking go off on full tangents based on little, and don't have the self-awareness to bring it back.
When you're a ChatGPT Projects or Claude Projects user, those caveats and provisos are your worst enemy because they'll change caveats into hard rules (either for the session or committed to memories) and you end up in absolute hell having to make it investigate to figure out why it can no longer produce anything but read-only pre-check code that never actually does anything but keeps performing stupid safety checks.
Lately it became load-bearingly-reality-difficult to not only read, but to comprehend the Claude output
I doubt they even intended it to be, but it seems like I kept going from resorting to 3.5-3.8 (over time) to realizing that Claude and GPT, while great at Python, will make rudimentary mistakes with R; even when they compose giant complicated R code.
Anecdotally I'd rate Gemini behind Claude and OpenAI models at fiction and I can't find any benchmarks showing Gemini is the clear winner at this task.
On my TODO is try and run all of the analysis pipeline in dense "machine speak" to save on tokens and just let Gemini sort it out at the end.
Only worked in a 1:1 in a quiet place. Still, can't complain for free.
> This mirrors how Apple has always segmented Pro vs. non-Pro iPhones: base models got LTPS panels while Pro models got LTPO, and only with the mainline iPhone 17/17 Plus did that gap close the standard versions previously lacked the smoother 120Hz ProMotion technology and the always-on display feature, unlike the Pro models — the 17e is the one model line still using the older, cheaper panel.
(emphasis mine)
I mean, I can guess what it is trying to say, but who RL'd this nonsense?
I've set my documentation sub agent to Gemini and my code agent to Luna
Ask yourself, how does Google -- a company that famously does everything -- benefit from SWEs outside of Google having access to powerful coding models? They would just be competition.
Famously, Google just eventually discards almost all businesses that don't have the same fire hose of revenue that ads does. Selling coding plans isn't something they are going to want to do.
Google is clearly motivated to make better search and information finding tools and stuff that will ultimately drive users through their existing search/ads/youtube ecosystem. That's really why they're in Android, that's why they do Chrome. Everything else with them is a sideshow.
Google is also full of beancounters obsessed with data centre quota and resourcing. Even massively profitable ads projects have to justify and fight for it. (Source: used to work there).
I can't think of anything less resource & revenue sensible than providing outside parties access to your TPUs for the purpose of letting them write stuff which could just end up competing with you.
Yes, maybe as part of their cloud business, selling token access could be useful money. But I doubt they'd tune it for coding.
Bragging rights to say they have a SOTA model. I guess that was more like Google of 10 years ago with moonshot projects. Nowadays, yeah, perhaps if it's not helping sell ads, it doesn't make sense.
Catch is, even Googlers internally do not have access to top-tier models. (or did not until recently, when apparently Claude was made accessible to the SWEs internally).
I think they’ve made a shrewd move in focusing on search integration and everyday users (Gemini app) vs software power users. They have their corner and nobody is really competing with them, plus it feeds directly into their existing revenue stream.
All audio generated by our AI products is watermarked with SynthID. This imperceptible watermark is woven directly into the audio output, ensuring AI-generated content remains detectable to help prevent misinformation. For details on our approach to safety and responsibility, review the model card.
As PrimeTime said, these are the guys that invented the 'T' in 'GPT', that deployed their first TPU in 2015, that is using hundreds of billions of dollar on AI - and they are beaten by 300 people startup named Moonshot AI even.
People are going to write books about this complete fumble.
My advice is to listen less to brainrot 'influencers' that optimise for engagement through sensationalism.
And they're making money doing it.
Perhaps they don't have the best coding model right now (although 3.8 flash is arguably SOTA at some benchmarks), but is that the be-all and end-all of AI? Only coding matters?
And Gemini is kinda good enough at everything. Never the top, but it is decent at every task, and it is much faster than Kimi K3 and significantly cheaper. Kimi is very focussed on coding, Gemini isn't.
More importantly, it natively understands text, audio and video. If/when we are able to make the jump to robotics, this becomes essential. As you say, Google has a lot of deep background and deep pockets, they are able to make more of a long play. No idea if it will pay off, but it is way to early in the game to count them out.
However, the "Extended Thinking" should be renamed to "Slightly Extended Thinking". Considering that it's the maximum thinking option for Gemini in the chat UI, it doesn't actually think a whole lot, leading to an uncomfortably high number of incorrect replies.
Nothing but constant errors with cryptic messages.
Let me tell you unlike every other mentioned model Gemini 3.8 Flash trial had to be reverted the same day. Instead of simply delegating tasks it would invent additional requirements and implementation details it knew nothing about and no amount of convincing not to do it would work. That's the first time a model failed on me so spectacularly despite having practically same Artificial Analysis Intelligence Index as another model that just worked (and higher than working DS Flash).
The reason I think it is relevant is: Live is likely even stupider model in every way possible (except hearing better than separate STT). So beware using it for agentic scenarios.
There's one or two I find more understated but I would love a 2026 SOTA V2V model that speaks clearly but without the artificial personality layered on.
Human interaction/theory of mind relies so much on non-verbal clues for interpreting emotion/intent and so for me having those neurons firing constantly while talking to an LLM just for an emotional no-op is exhausting to put up with for more than a couple minutes.
There's one male voice that would make me assume someone was sarcastically mocking me if I was talking to an actual person because it's just so over the top.
I don’t want to be aroused by my turn-by-turn street directions, thanks.
Copes well with thick accent, voices are pleasant and latency seems low.
Oh and I can actually use it on a workspace account - which for most of the recent releases was an account stuck in limbo. Not personal enough for personal offering, not enterprise enough for enterprise.
Well done G - will definitely be using this
Is this a pure TPU infra? Really high performance solid intelligence.
I've been using Gemini to live chat in Afrikaans and do impromptu Afrikaans grammar lessons during my solo drives around town. It is phenomenal at speaking the language - like, it really shocks my family members when they hear it.
This is probably the most joy I get from any of my usages of LLMs/AIs. It's been really, really nice getting to speak my language regularly again. =)
So, I'm excited about this release and live chat getting better. I also hope the other frontier labs pick up niche languages like this as well so that I have more options.
Looking forward to where this can go.