At this point if a Google PM handed me $100 and said "go figure out how to issue a prompt to Gemini 4 Argon" I'm only ~30% sure I could figure out how. The Gemini app in my phone only has 3.8 Pro as the latest model; it does say "Ultra" in the bottom left, but I really don't think I'm on the Ultra plan. How I would sign up for that, on a Workspace account, is beyond our brightest minds.
Its sad watching them try to keep up when they want nothing more than to crumble under the weight of their own bureaucracy, and at this point I'm not sure there's any hope for them.
This model is not for you or me. This model is for Google AI engineers to put on their perf review at the end of the year. How you find or not find it, it doesn't matter. It goes into the "designed and developed Gemini 4 Carbon Ultra Pro" paragraph and people get a promotion, get to keep their jobs and hope to not get replaced by Gemini 4 Carbon next year.
That seems like an overly negative caricature of Google’s corporate culture, and unlikely to be the case here, given the stakes and enormous investments in AI.
It’s more likely what the grandparent was saying, I’m reminded of Steve Yegge’s rant from 15 years ago and it seems like not a lot has changed.
What gets me about that email chain is that Gates complains, but doesn't provide any guidance, instead leaving it up to the people who make the mess to try and fix the mess.
Compare to leadership likes Jobs, where he would spell out how it needs to function.
Yeah - I'm not too sure what the discrepancy is either. "crumble under the weight of their own bureaucracy" and "How you find or not find it, it doesn't matter [because the bureaucracy incentivizes self-promotion and not user impact]" seem to be more in agreement with each other than disagreement.
"overly"? surely not... maybe in the sense that it isn't "underly", but there are much more uncharitable caricatures of Google's corporate culture that I would freely admit than that.
They don't give a shit. You can use Argon on Pro but not that pro and no not on that app which we depreciated but on this new app which will be depreciate before you can install it.
It honestly is really hard to figure out what to use and where. It shouldn't be. A pro plan is a pro plan. It doesn't matter if you get it via x,y or z. If there is difference based on where you get it, then call it something else and not a pro plan.
They make something are feel like its the best thing in the world and leave users behind to experience it hiding it behind obscure docs and hiddens T&C with no support. And users have no clue what and how to do it.
I wish this was a joke. Turns out, Google turned into a joke. I was really rooting for them; They had everything in place to be the top dog. From people to research to hardware to internet.zip (and video.zip and emails.zip and newsgroups.zip and code.zip and android.zip etc..). I don't want to sound negative and what they built is probably great in its own right, but what was expected of them and what they have delivered so far, I can definitely say that they blew it hard. They did manage to turn gemini into search 2.0 on google.com and it does work.. but, come on. Where's the google of yesterday? Seems they put a ton of weight that's not on the level of people that were there before and I don't want to offend anyone there, but it sure looks like that.
The same was true 15 years ago when the cloud started hyping up and Google missed that in similar ways as well, so I’m not sure whether the “Google of yesterday” is who you think it is.
This is the Google of yesterday, they didn’t change.
Google finance was an incredible portfolio tracking and research tool. Until they released v2 which took away all the functionality and replaced it with a news feed.
The comment you replied to was: "The only good products Google has ever built in-house are [...]"
I don't have enough information to judge how much of Chrome was original work and how much was pre-existing work. Though I guess by now, after nearly 20 years, Chrome's codebase looks absolutely nothing like KHTML.
It was the rare confluence of people needing the same thing the megacorp showing them ads needed. We all benefited, no hard feelings, but always remember to look at where the puck is going.
The fact that it made the browser itself impossible to monetize to fund development was a part of it. They were an antitrust insurance to Google and tried lots of stuff to improve its position. Granted they failed, sometimes quite spectacularly, but I still drive Firefox for personal use.
The money earned with Firefox should have gone to Firefox but went to vanity projects instead. Firefox OS, Mozilla Social, Firefox Send / Persona, acquiring Pocket, etc etc. And all of it failed.
Nevermind the controversies of silly ad campaigns and CEO salaries and whatnot.
Mozilla should have just reinvested all the money into Firefox. Heck, they couldn’t even finish their rust-based rendering engine.
They’re still the same price as an uber and there’s no way to actually order one, it’s all random. I don’t understand why… like I actively would prefer them. If they were available so instead I find myself reaching for the cybercabs instead at half the price.
Google acquired it from Upstartle (the company behind the online word processor Writely).
Writely was created in 2005 by Claudia Carpenter, Steve Newman, and Sam Schillace. Google bought Upstartle in 2006 and Writely became the foundation of Google Docs. Google then combined it with other acquisitions, such as 2Web Technologies for spreadsheets, and launched Google Docs & Spreadsheets in October 2006.
App Engine was good for its time, and I don't think that was an acquisition. Similarly, Cloud Storage and Google Cloud in general are pretty good products.
I started using AWS when EC2 and S3 first came out, but was forced to switch to Gcloud by the most recent company I joined. I was pleasantly surprised. In many ways it's more techie-focused than AWS, which is more enterprise-focused. And their managed Kubernetes is second to none.
You want an executive to issue a press release about how dysfunctional and bureaucratic their company is and that’s why it’s so hard for users to actually use these new models?
Your sarcasm is as low-effort as your original post.
Anyone with a functioning brain can see how
> You want an executive to issue a press release about how dysfunctional and bureaucratic their company is and that’s why it’s so hard for users to actually use these new models?
...is intentionally assigning beliefs to me purely from your own imagination without any connection to reality.
Your childish pretending that you didn't say something that is literally on record immediately above is a tactic employed by kindergartners, caught by their parents eating a cookie, who claim that they're not eating a cookie.
> Your sarcasm is as low-effort as your original post.
Putting too much effort into sarcasm doesn't work well usually; it works because it's low effort.
> is intentionally assigning beliefs to me ...Anyone with a functioning brain can see ...caught by their parents eating a cookie, who claim that they're not eating a cookie...
Who is assigning what beliefs to you? What cookies? Not sure what you're after here but sounds like you're a bit upset.
> Putting too much effort into sarcasm doesn't work well usually; it works because it's low effort.
Your comment is exceptionally low effort by the standards of sarcasm. llama-3-8b-level, in that it's trivial, uninteresting, unamusing, uninformative, not clever, and incorrect.
> Who is assigning what beliefs to you?
You did, in the comment that I quoted from you immediately before my statement.
> Not sure what you're after here but sounds like you're a bit upset.
We've established that you have either a severe lack of reading comprehension or are maliciously lying, given that someone with rather below average intelligence would be able to tell that your comment assigns a belief to me that I never held, by reading the two sentences immediately adjacent to each other.
Most HN readers are capable of this reading comprehension, so anyone reading this thread will be able to tell that you're wrong, and I can link to this in the future if you make similarly deceptive comments.
> Your comment is exceptionally low effort by the standards of sarcasm. llama-3-8b-level, in that it's trivial, uninteresting, unamusing, uninformative, not clever, and incorrect.
Ok. Well, I found it amusing and interesting.
> You did, in the comment that I quoted from you immediately before my statement.
You still haven't quoted them. Just vague references and a lot of angry words.
> We've established that you have either a severe lack of reading comprehension or are maliciously lying, given that someone with rather below average intelligence would be able to tell that your comment assigns a belief to me that I never held, by reading the two sentences immediately adjacent to each other.
You still haven't quoted anything are just continuing with more insults.
> Most HN readers are capable of this reading comprehension, so anyone reading this thread will be able to tell that you're wrong,
Yes they are and it would look rather embarrassing for some of us... but I can't stop you ;-)
> You want an executive to issue a press release about how dysfunctional and bureaucratic their company is and that’s why it’s so hard for users to actually use these new models?
There we go. That's the quotation. Your claim that I never quoted is objectively, factually, trivially false.
This is wild. You literally not only forgot your own words that you wrote less than 12 hours ago, but you were incapable of reading the comment where I very clearly cited them right back to you.
This is sub-grade-school levels of memory and reading comprehension. That's not even an exaggeration - grade schoolers are easily capable of doing the ctrl-f levels of "find this exact string of text in this other string of text" that you just failed.
>> You want an executive to issue a press release about how dysfunctional and bureaucratic their company is and that’s why it’s so hard for users to actually use these new models?
> is intentionally assigning beliefs to me purely from your own imagination without any connection to reality.
No, that part is asking a question. I assumed you misquoted that part and meant something else. Nobody is assigning you beliefs. The normal response to someone asking a question is saying something like "yes, that is exactly the proof I would expect" or "no, I would like an insider to come forward and give an account of X,Y,Z". You did neither and instead started with insults
> This is sub-grade-school levels of memory and reading comprehension.
There was nothing about my comment that remotely motivated a question like the one you asked - but you were, quite knowingly, implying that I believed that. This is known as a loaded question, and it's unambiguously malicious.
When I link this comment in other threads in the future, you'll be able to see that others can see it for what it is. Even on HN, only a vanishingly small fraction of individuals can't recognize a loaded question when they see one.
Edit: And, because you're acting in bad faith and will likely claim that there's no loaded question, I asked Gemini, Claude, ChatGPT, and Grok whether there was a loaded question in our exchange, and every single one of them said said that you embedded a loaded question (with ChatGPT also correctly identifying the implied accusation bundled in). Full prompt below - future HN readers, feel free to reproduce yourself :)
-------
is there a loaded question in this exchange?
A: This model is not for you or me. This model is for Google AI engineers to put on their perf review at the end of the year. How you find or not find it, it doesn't matter. It goes into the "designed and developed Gemini 4 Carbon Ultra Pro" paragraph and people get a promotion, get to keep their jobs and hope to not get replaced by Gemini 4 Carbon next year.
B: Uh, proof needed for this outlandish claim.
A: You want an executive to issue a press release about how dysfunctional and bureaucratic their company is and that’s why it’s so hard for users to actually use these new models?
I remember a year ago and a bit ago trying to get an api key for some Gemini ~3 model. I probably spent an hour trying to find how to get a key in AI Studio. Finally figure it out and got a key.
I needed batch mode. Oh but the key from service X doesn't support batch mode, you need to go service Y and btw there is no UI for that so you need to use an SDK or API (cant even remember anymore). Spend an hour figuring that out, then it turns out I can't use batch mode because I don't have permissions. Have boss spend an hour granting me permissions. Then service is unavailable every n requests for 5 minutes, so I have fuck with exponential back offs.
Glad to see they moved that experience to their consumer offerings lol.
Yeah, their native API reliability is absolutely horrible. Constant request failures. I've always had to go through Vertex or OpenRouter for any sort of heavy load.
This 100% being in the compute/ai console (i dont even know what is what) is easier to "google it" and hopefully land on the correct page. Same with AI billing. Can never find the damn page !
Oh and good luck figuring out model-api-names (same complaint higher up)
Technically not true any more (for some versions of copilot chat, at least) -- my workplace's version has a "GPT 5.6 Sol" selector with "Answer faster" or "Think Deeper". ("Think deeper" appears to represent the lowest reasoning effort).
But you have to select it each time; otherwise "Auto" sends you to a lobomised model that is probably Phi-4 :(
One of the unexpectedly useful aspects of Simon Willison's pelican benchmark is that you can estimate what model you're using via the pelican it generates:
"GPT 5.6 Sol": appears genuine, albeit low reasoning.
-> handed off agentically: appears to be using Luna for agentic tasks.
"Auto": very low quality, does not appear to be an OAI model.
Yes, but last time I asked copilot to provide an alt text for an image depicting a experimental procedure using rodents, it suggested "A rat being attacked by another group of rats".
It is part of Office 365 ( or whatever they are calling it now ) subscription and apart from coding, it is literally forced into everyone's throat in Fortune 500 for day to day task.
The problem is when they stop releasing big models for so long that their small models overtake large ones.
User wise, the average person will see pro and expect it to be better. Where as companies like OpenAI have tended to release across small/mid/large such that the disparity hasn’t been large in terms of a smaller model outperforming their latest mid/large model.
Not sure how much it matters for the average user though.
For myself and most others that actively use LLMs, I’m aware of the model names and differences, and on the off chance I don’t, I can do a quick search to find out.
On the other hand, my dad (“average user”) probably doesn’t care whether to use Argon or Pro, or Flash for that matter. They are all good enough for his use case anyway.
I suppose the “Auto” option is what AI companies expect the average user to use.
Not sure I agree with this argument, it's basically saying that it doesn't matter if stuff is badly labelled as people don't care what they're using anyway. fanta can be labelled as coke and coke can be labelled as fanta because the person consuming it doesn't care if they're consuming fanta or coke.
I think in many cases people may not be very AI pilled / be looking at artificial analysis etc and might think "I'm doing this much harder task, let me use the best model" and they may infer from naming that 3.1 Pro is better than the recent flash models which would be incorrect.
Anthropic had an issue with delaying the release of a new Haiku for ages but it didn't really mess with things as it was always the weakest model and so positionally it stayed the same. If however they stopped releasing new Opus models and somehow Haiku started overtaking Opus I think it'd confuse the normie consumer.
I was confused a bit where this was.
Then I had to Google Gemini, which lead me to gemini.google.com (did not know it existed, and I use gemini in prod via google cloud).
First thing there that got me was that all of my Google searches. Why?
I googled "snoop dog real name". Why is this considered a gemini chat now? If I actually used it, my chats would be buried under dumb quick google searches I did. And it seems its only my phone searches, not alL?
I assume because some PM will say "we have 1243153344% more Gemini searches than last quarter", then gonna get paid 3 million USD bonus for something that ruined the UX.
Then, I looked at the top and didnt find the model dropdown.
It is next to the input, okay. I click it.
Argon? Locked. Even tho I pay for the API anyways and can use it via the API.
But lets continue.
My app stopped working 2 days ago for no reason - apparently, I needed to move to a "prepaid payment mode", which wasn't a problem for the last 2 years.
- I open the AI studio, the banner to update shows, then page reloads and it disappeared.
- I was like okay, lets see - checked billing, tried to top up. First time my bank's 2FA notification didnt come, so I rejected and tried to top up. Then it would reuse the same windows as before with rejections. Okay, I refresh, topup, it works. I try to open usage to see my usage, but I can't see the spend there in any currency besides tokens.
- Now let me try it out. I type in the prompt, select the model, hit enter. Oh wait, I have to link the API key. I click the button, a dialog opens, then closes itself without any information. The prompt? There is no prompt anymore, you have to type it out yourself.
While I believe in Google's AI offering - I use their models and for my niche they are the best for good quality at great speeds.
But holy shit, every time I use one of their products, I just get annoyed.
While Logan and AI studio team have really done some big improvements, the overall Google offering is confusing as hell.
Why does a customer have to hear "Oh yeah, don't use Google Cloud now, use Vertex via Gemini Enterprise Agent Platform"? Because some PM chased a bonus?
The internal politics are definitely ruining any edge they might have.
> First thing there that got me was that all of my Google searches. Why? I googled "snoop dog real name". Why is this considered a gemini chat now?
It isn't - it's probably an _assistant_ query? In fact the Gemini and Search orgs are cartoonishly at loggerheads; both developing more or less the same product (http://gemini.google.com vs http://google.ai).
No, it is a search from my launcher on S24. Literally google searched from the launcher as I always do. I specifically remembered that, because I was sitting on the couch with my GF watching a movie and quickly googled his real name.
Now that I look at it, weirdly, both show different results.
Gemini shows only what I searched on my phone, meanwhile Google.ai shows ALL of my searches as chats.
Guess both PM's are getting a big bonus this year, and the stock price will get a nice bump from "Look at how much people use our AI".
Weird, maybe it's specific to your launcher? My searches from the -1 page of the launcher don't appear in Gemini, and the ones from the bottom corner swipe to search or OK Google don't appear in AI mode.
Didn't for me (antigravity, pro sub). I got it working with the API, but to my dismay I learned that counts as enterprise and so many features are disabled, such as auto mode.
I only use Pro for Google as their models were so far behind, but I tried to give Argon a go, and notably it is only available in the chat interface. One brief conversation, with three exchanges, and my quota was exhausted and it restricted me to Flash-lite for the rest of the day. Rather comical stuff.
And it prompts me to upgrade to Ultra for "5x more usage". In contrast I use my Claude Max plan endlessly, on enormous problems, and my 5x max plan sees me almost never hitting double digits on my quota using Opus 5.5. It doesn't compel me to sign up for Google's higher level plans.
i hear this but i can't believe it. I'm writing multiple moderate apps. rebuilding entire distributions. not coming anywhere near hitting the limit. Pro plan (on a discount too). What am I doing wrong (right?)?
Im using pro on a discount. I got argon a few days after the release but the usage limits are worthless. Asking it for a simple review of a blog post blows through 50% of the usage limit. And in the limited experimenting I have been able to do because of that it also isn’t smarter than 3.8 flash.
Could it be skills / customizations? There are tools that will help you understand where your tokens are getting burned.
Making sure you are scoping your context, being specific also helps.
And I do 90% of my work on 3.8 - flash - low. C, go , JS, bash , powershell generated content. Near perfect quality.
I hear lots of people say the budgets are low – I’m absolutely hammering it sometimes indiscriminately (shell into this instance, rebuild the entire vm image).
I’m using the default anti gravity setup. 3.8 medium is my daily driver. This gets me pretty far. (Definitely not as far as Luna and Sol). When I ask the same question to Argon it is 50% of my weekly usage limit gone. So either the usage limit is really a lot smaller or somehow argon thinks a whole lot more even for simple tasks.
i’ll have to test that out. I’ve used argon a bit, but i guess not enough to really push the quota. The actual token based pricing is only 2x
What tasks have you needed it for? So far I’ve only switched for experimentation and while it was better, it didn’t seem necessary, especially given the latency.
I purely intended to use it for review and architectural design given my bad experiences with Gemini coding. The bad experience stemmed from it getting stuck in loops and burning quota. The code quality when it worked was great.
I now asked Argon for a pretty elaborate review and it used only 2% so I will be giving it another shot.
Another thing that can be of influence is that for subscriptions, Google uses some proprietary idea of “work done” for the quota instead of simply how many tokens were burned. I can imagine that this heuristic is sometimes quite off.
Using Argon? Are you certain you aren't using the Flash models?
Their Argon limits are the real issue, Flash is ok, but honestly compared to Clause's limits they are all horrid, OpenAI is the next best option if you want Astra, otherwise suck it up and send money to DeepSeek or find a provider for it for cheap.
I'm on the Pro as well, haven't checked if argon became available on antigravity, will tomorrow, thanks!
I noticed the usage on chat is absymal, but it could very well be that their own harness does a good job in caching and you can squeeze good usage out of it due to this.
It's at least not worse than Microsoft. Go and start a chat with Copilot. Which one? I don't know, they are all the same. Or maybe they are all different? All I know that if you start a chat in one interface, there is close to a 0% chance (but not quite) that you will find the same chat in another interface.
I had Claude write me a script to probe all the GCP Vertex models. I can say with confidence that I know how to call Argon, because Claude told me how, I did it, saw it was modestly better than Flash for my use case and I decided it wasn't worth the money right now, for this use case.
I'm fairly confident I can't call Carbon, and I am okay with that.
>I had Claude write me a script to probe all the GCP Vertex models. I can say with confidence that I know how to call Argon, because Claude told me how,
Wow that's embarrassing ! Is there no "product owners" or people "fighting for the users" anymore at Google ?
Unironically it's easier through third parties. Cursor has had Argon available, but in our enterprise we cannot get argon from the gemini subscription because they have some weird license split going on.
I don't even read Google's release notes anymore. What's the point? The models are never available.
Now, I just go straight to the Hacker News comments. I've given up on trying to figure out how to navigate Google's labyrinth.
Their release notes have become like breakthroughs in graphene. I'll believe it when I see it.
Their go-to-market execution is just so broken.
At the beginning of this whole AI shift, when Google would announce a new model, I would try to get access, even paying for a subscription.
Then I realized it's just not possible.
Google onboarding is the new Gordian knot.
For Google, I think the announcement is the success criteria. And onboarding is somebody else's problem. And the problem is... that somebody doesn't work at Google, or are entirely disempowered.
If, someday, Hacker News comments indicate that a majority of individuals have figured out how to access these newly announced models consistently, I might take my precious personal time to navigate the labyrinth once more.
All I see is 3.1 pro and 3.6 flash, I just canceled the sub lol - those seem like half a year old and I'm too lazy to check - my workplace gives me access to all models in a harness anyways
Tried to add AI Ultra or Pro to my business workspace account and I wasn't able to do that a few months back, then tried again when Gemini 3.8 news came out, failed again, then finally upgraded my personal account to AI Pro, tried gemini 3.8 flash on my projects, it was worst model ever. The only benifit I got by having AI Pro is that you can use Youtube adsfree.
I too could not figure it out. Also a shame the Vertex API format is sorta different to OpenAI's, which everyone but Anthropic uses. The sanest way to use Google AI models is via OpenRouter tbh.
Google's UI is always bafflingly overcomplicated and opaque. I never understand why such a large bunch of very intelligent people cannot create a useful UI
I’m with you on this. I tried to get Argon access for our org and failed. gave up. moved on. Still just 3.8 in the dropdown for us.. Thinking its just in so Ai studio app or something and not in gemini app?
Why is it so hard for a decent PM, Designer, and Eng team to build out a better version of Antigravtity with an equivalent web interface for general chat? I don't get it.
Same here. When I use the Gemini app for iOS, it says that I’m on “Ultra” in the left sidebar near the bottom. It doesn’t say that when I’m using Gemini on the web.
the charitable explanation is that they derive the most important revenue in this area from enterprise. And enterprises are not allergic to horrible product experiences and upsells.
The more likely explanation (even if the above is technically true) is that Google is not good at cohesive product experiences.
Not disagreeing with the premise - I've fought with this in the past - but I'm using Antigravity CLI and Antigravity desktop (GUI), and both automatically started defaulting to Argon 4 as their model a while back.
I have an AI Pro account - which I decided was a no-brainer because of the 5TB of Gmail / GDrive / GPhotos space it gives - and this all just comes along with that.
One thing to check, though, is that the client is running the latest version. At one point my Antigravity GUI, which had been working fine, started listing only Claude and GPT-OSS models. It turned out to be because the GUI needed to be updated. I guess the Google API endpoint had changed behavior without any backward compatibility.
So yeah, they're hilariously bad at rollouts, but it's not that difficult if you're technically inclined or ask an LLM.
Benchmarks are just benchmarks and they could have been carefully selected but it's not "trying to keep up": they're ahead of Fable 5.1 and GTP 6 Astra in 15 of the 17 benchmarks. And in many cases by a huge margin.
Instead of talking about that, let's talk about how difficult it is to pick the model or how crazy Google's product naming methodology is.
Not saying they're not concerns: but let's not downplay what the model is achieving by saying they're "trying to pick up" because switching model ain't as easy as with the competition.
I mean: I don't use a model based on how easy it is to switch to another model. That's not what I'm looking for.
I don't think the models being a commodity matters much. Consider: Email providers are a commodity. Cloud linux server hosting is a commodity. Checking accounts are a commodity. There's a lot of great businesses, trillion dollar businesses, built on commodities. DeepSeek v7 Pro might open source and be better than any Claude model (unlikely), but JP Morgan and Eli Lilly are not going to go spend $500M on chinese AI tokens routed through OpenCode. They just aren't. The models really matter far, far less than the experience these companies build around the model, and both Anthropic and OpenAI are doing a decent job of making that experience reasonably sticky and compelling.
"trillion dollar businesses, built on commodities"
Which ones? I'm looking at the list of companies with > $1T market cap and none of them have commoditized markets. Instead, they all have huge network effects (Apple), enormous legacy lockin (Broadcom), a religious monarchy (...), or are some of the only people on the planet who can do what they do (nvidia, tsmc).
Mind you, there's lots of _billion_ dollar businesses in commoditized or partially commoditized markets. But Anthropic as a billion dollar business is sunk.
> DeepSeek v7 Pro might open source and be better than any Claude model (unlikely), but JP Morgan and Eli Lilly are not going to go spend $500M on chinese AI tokens routed through OpenCode.
But they might spend them on American inference providers running DeepSeek v7 Pro with ZDR routed through OpenRouter. Or hell, they might just run it themselves like we're already seeing some enterprises do right now.
They might, but I am skeptical of this idea. The story of the past fifteen years in enterprises has been pretty clear: Move things to the cloud, we don't want to manage data centers. There are notable exceptions, but by-and-large that's the trend, and I don't understand why people believe this will suddenly reverse in the era of AI. Running a GPU cluster is substantially harder and more expensive than any other cloud resource. These companies don't even want to run their own Microsoft Exchange servers anymore, and you think they'll want to run an NVL72?
To be clear, I don't like this trend, and I strongly believe most companies are better off owning as much of their technology stack as possible, working with partners to help maintain it rather than making themselves lesser by working with AWS/GCP/Azure; but it is the trend.
Something like Azure Foundry w/ ZDR is more likely. And yeah, its possible that Ant will find competition from those kinds of providers. But Ant will also find revenue from those providers, and its hard to say where things will land in the end.
You should go look up the weight of a 911, Boxter, or Miata. And then, I suspect, you will not return here to reply to this comment. You'll invent some other reason why EVs can't be "nimble" or whatever, because it can't be your worldview that's wrong, it has to be the world.
Have you actually taken any of these cars along a twisty country road? They are not at all similar.
911 is not really a roadster - it's a full on sports car. Even then the older ones top out at about a ton. Same as the Boxster and Miata.
Most serious sports EVs start at 2 tons or more. It's really impressive what they can do to hide the weight, but because it's all digitally controlled, it's really not a dynamic driving experience in the same way.
More importantly to my point - all of them have nearly identical drive characteristics. Outside of raw performance, if you put a gun to my head and asked me to describe the difference in handling between a Tesla, Mach-E, or a Polestar - I'm not sure I could give you anything.
Oh, it doesn't require a person to continually monitor it now? I missed that change. Until that happens I don't think it qualifies as "full" self driving.
Only from a regulatory perspective. Anyone who drives a Tesla for long enough knows how to defeat the monitoring system, and doing so is extremely safe; significantly safer than a human driver. I would have no concerns summoning it autonomously over 100 miles, once the law allows for that to happen.
In a single 5 mile trip within the last month, an FSD Tesla I was in made four mistakes that were saved by the actions of other drivers on the road. I have concerns about you summoning it autonomously over 100 miles.
Well, you're insane for thinking that. Drunk driving is very unsafe, and population-level statistics don't lie. They're quite conclusive when it comes to grading the safety of drunk driving as being dangerous relative to sober driving, and Tesla's full self driving system as being safe relative to sober driving.
Teslas already have very strict (annoyingly strict) gating keeping specific functionality off of public roads (Summon). The drag mode thruster-assisted boost will probably leverage that. I also imagine that cameras on the car would stop it from being engaged if it detects humans in its firing area even when not on public roads.
The flying trim will never be sold to general consumers. Even the rear-thruster trim is currently listed at a $250,000 pre-order fee, which is a deposit for a fraction of the full cost of the car. Final cost will be over $1M. Any of the thruster variants will be extremely rare; like, the one guy in your city that has one will show it off at car shows twice a year, McLaren P1, level of rare.
Yeah they'll make a few dozen of them and sell them to people like Joe Rogan. There is zero chance that any individual who doesn't have the personal phone number of a Tesla leadership team member will be able to buy one of these.
Generally, very true in mental health, but moderately true for all areas of health: The idea “at least we have systems/doctors/drugs to help solve this if/when it becomes a problem” is a massive cognitohazard. You do not want to be on the other side of that fence, and belief in solutions there will slide you toward it quicker. Do everything you can to avoid it while you’re on the other side.
e.g. if you aren’t getting 10,000 steps every day for six months straight, you aren’t doing everything you can. It might not solve your problems, but you don’t know that, because you’ve never tried. If you aren’t quitting nicotine & alcohol, you aren’t doing everything you can. If you aren’t saying hi and being shot down by at least one stranger you find attractive every week, you aren’t doing everything you can.
Decision models have the potential to have an even larger impact on the Real World than LLMs have to this point (which is obviously quite large). But the model itself matters less than the product experiences you build around the model, and its very likely that the incumbent labs are treating the area as something more like "oh yeah I guess we can ship that and then forget about it" rather than investing in what building business processes on decision models looks like. Unlike full language models, I don't think the primary business of Typesafe will be serving Jev at API pricing; it'll look a lot more like putting Jev at the center of a much more expensive suite of software.
There's the potential for an inverse LLM play. In contrast with LLMs, all that seems to matter is the model, and the products the labs build around the models are all really samey and boring; the same left panel list of agents, main view agent conversation, right hand extra context, and we're now in the era of everyone creating the same cutesey furry friend on top of all this tech.
> But the model itself matters less than the product experiences you build around the model
This is a very important insight. And it applies to LLMs as well. Very few people were impressed with the capabilities of GPT 3, it was mostly a techie novelty
But then when they added chat on top of gpt 3.5, all of a sudden it was a huge hit. Sure there were improvements in the model from 3 to 3.5, but the biggest impact was from the chat experience
Conversely, when they created Eliza, a basic chatbot more than 50 years ago, people even got addicted to it, despite it’s ai model being something super rudimentary and basic compared to what we have now. The model capabilities didn’t matter as much as the experience the chat created
The instruction tuning between 3 and 3.5 was a big deal. If you prompted raw GPT 3 "Tell me a story about a mouse", it was as likely to continue "2) Tell me a story about a cow. 3) Tell me a story about a rabbit" as it was to tell you a story. I tried to use GPT 3 for work and found its only utility was to get over blank page fear by writing something so bad I would fix it in anger.
It was impressive, but not that much. If you had a basic prompt prefix like:
“The following is a conversation between a human and a helpful AI assistant.
Human: [prompt]\n
AI:”
You would get consistent conversational chat, with the main limit being the model’s limited context window, and obviously frontier intelligence at the time.
I played around quite a bit with davinci-003 and conversational systems before ChatGPT. I kept using this format and API for a while, but at least during the ‘free research preview’ era, found it generally smarter than chatgpt. 3.5-turbo was probably the watershed moment where I moved away from completions on base models; to chat completions.
If you’d like to emulate this experience, spin up a base (not instruction tuned) checkpoint of Llama 1; or a more recent base model. You might underestimate how much ‘intelligence’ you can get :)
Back then the API exposed a lot of controls from sampling to logits, so you could specify a custom stop token (some rarely used Unicode); then switch to near-greedy sampling for more reliable “function calls”, etc; and then switch back to sampling for chat.
In the early days, before releasing something with the API, you had to get your use case approved by OpenAI, like the App Store review process.
I remember those days and I'm glad to see another one who's been in the industry since the very beginning :) I was quite sad when davinci-003 was taken away; still remember the green-yellow-orange-red highlights the OpenAI console would make for log probabilities of tokens and all the cool things we could do with a pretrained model.
In my courses, I teach students about that era by using gemma 4 (pretrained) and showing how you can emulate a conversation even on something that wasn't post-trained for chat.
I think the main benefit of instruction finetuning was that the model is less likely to get prompt injections and more likely to follow the specified format?
Pure davinci was easily confused by what was system prompt, user prompt and its own answers.
To be fair, there are still many jailbreaks for instruction tuned models. Surely less than before, but it could be because models got smarter (Claude often "reasons" that something is fishy even if it appears in the prompt or appears like a system prompt).
> Decision models have the potential to have an even larger impact on the Real World than LLMs have to this point (which is obviously quite large). But the model itself matters less than the product experiences you build around the model, and its very likely that the incumbent labs are treating the area as something more like "oh yeah I guess we can ship that and then forget about it" rather than investing in what building business processes on decision models looks like.
I agree that the potential is enormous, but I'm not sure why it would be Typesafe who figures out the best way to develop and package this capability. They have a bit of an advantage in the amount of time they have been thinking about this problem specifically, but I think it's a big leap to say that they will be able to dominate the category.
A major limitation of directly using "System 1 decision models" (a.k.a., logistic regression, and related uncalibrated classifiers over the output/logit space) in enterprise settings (or other high-stakes settings) is that such estimators are not reliable estimators of the predictive uncertainty in the presence of covariate shifts, and such estimators also lack a means of instance-wise data attribution (i.e., interpretability-by-exemplar), so they're not the ideal estimator for verification, routing, uncertainty over retrieval and tool-calls, etc.
For decision-making with neural networks, we instead need the older idea of estimators of the predictive uncertainty with constraints in the feature-representation space (over training/support of the estimator), as with Similarity-Distance-Magnitude estimators: https://pypi.org/project/reexpress-sdm/
If youre the one or have spoken to someone implementing "AI solutions" inside large companies recently, a decent chunk of it is soft policy enforcement with very basic context. They moved from gemini 2.5 flash lite type models to jev. Which is also why I found the price comparisons to "GPT Astra" on Twitter rather funny.
Right, example being expense pre-approvals (for low dollar value purchases like office equipment). General purpose LLM not required, and the product experience surrounding it does matter (reporting, auditing, fine tuning future responses, etc.).
It really shines when you use that decision routing in industries where they need to automate these routings reliably with very low latency, higer certainly and ,more important, at scale. Take call centers for example. If a voice company needs to automate escalation etc. an llm would be an overkill. That's where system 1 models shine.
If it was simple they would just use an LLM to write code to do it. These decisions are “fuzzy” in the sense that they are difficult to make in just code. They need the “smarts” of a language model to evaluate something, but the final answer is basically trivial (e.g. choose one of these five options, not write me 20 page of text or one shot this whole complex program).
There are a lot of those types of problems in the world, replacing what was previously spotty hardcoded decision logic that worked 70% of the time with something that now gets it right 95% of the time. Even if not perfect, the fact that you can dramatically close the gap means a lot lower cost overall, so it’s worth it.
It is. But the thing is, that is also a lot of the usage for "AI" these days. LLM based search pipelines are being deployed but not really used. Chatbots are a great hit. And then of course, in coding, finance, sales, accounts and audit, etc, it is being used to accelerate output.
These guys also love text to sql. I know 4 people each implementing text to sql for their separate companies. Or text-to-redash dashboard in some cases. Funnily enough, text to sql is one of the things where LLMs are clearly sub-human. Mostly due to lack of easy verifiability.
> These guys also love text to sql. I know 4 people each implementing text to sql for their separate companies.
Text to SQL is a really bad idea, as the agent will just generate really convincing and terrible analyses that lack all business context.
I mean, you could implement some kind of semantic layer, but that's a lot of work, for uncertain reward.
I saw a place which essentially tasked the data analysts/scientists in an area to describe the good tables, and assumptions and necessary context to make decent queries/analyses, and that definitely lead to much, much better (almost usable) results.
As I have said many times, what the world really needs is SQL2Text, rather than Text2SQL.
> I saw a place which essentially tasked the data analysts/scientists in an area to describe the good tables, and assumptions and necessary context to make decent queries/analyses, and that definitely lead to much, much better (almost usable) results.
Yep. After this has been done, the later text to sql outputs works half decent. The major issue I heard was that whenever a small assumption changes few months down the line, while humans are able to "JIT" apply it to each of their queries, these models sometimes lose it in all of the context and so it becomes hard to progressively update anything - you end up having to redo the whole workflow.
And in my experience, even the more complex dashboards requested by product or upper mgmt were stood up within the day by humans. Or the next day (but that was cos some ETL would be scheduled overnight rather than immediately). And I am not sure if speeding up this part would even matter to the guy who requested the dashboard. So I am not sure if this is a major bottleneck people should try to innovate in. If openai does the work for you then great but otherwise. Humans aided by copilot-of-old style autocomplete is probably more than enough.
Needless to say, text to sql is pointless for application queries made during API calls.
Yeah, it's not the development but the maintenance that kills you in both software and other information fields.
The whole text to SQL thing is about data teams being a bottleneck to the business (rather like software people). The goal is good in theory but honestly an Excel plugin is probably a better solution in many cases.
It's pretty high value add when you consider it's replacing human beings assessing each case. This category of thing is present all over enterprises. It was outsourced to BPO shops in India and the Philippines before, now it's going to AI.
But you're right that more complex solutions are still a bit elusive. I use a bot to plan my strength workouts based on my soreness level and what equipment I have available and it's decent, but not perfect.
The simplest example could be replacement or at the least an economical complement to the LLM step in cascading VoiceAI, almost every start of the turn a local JEV could make a decision for end of user turn detection:
- is the user finished with their thought? If so delegate to next system(LLM another JEV)
- Should I wait for the user to complete their thought for a few seconds? Set a timer and when time out expires, ask the user to continue.
Like just tuning this model can make it cheaper for voice AI providers to do semantic turn detection over the space of responses.
That was my first insight as well. i have been working on a voice agent for a while and decided to spend about 4 days messing with system one models to see where it could fit in my cascaded system. after 4 days i ripped it out of my system as i found better solutions.. i tried really hard to find a use for it but every time i found there was no need for it so i stopped trying to shoehorn it in. i also spent time with the vision side of things for this thing, and i think that where the biggest gains will be IMO. making decision based on visual ques and in pixel space only was the biggest win.
Yes, i looked in to pipecat stuff a while ago and revisit it often because its relevant for voice agents, but i never ended up using any of their stuff as my custom solutions were better. Also id like to hint that smart end of turn detection is not needed at all in a cascaded system IF you have access to the inference engine of the llm you are using in the sandwich. I dont want to reveal it all but lets just say its no longer a headache or any issue at all. If using 3d party api then yes it helps to have a smart eot system. But i try to keep everything fully local.
Yeah agreed that vision is a nice fit here. I’ve gotten some really impressive results by just throwing xgboost on 25 classification related questions.
System one models do not have an obvious use in voice ai, unless you are choosing from a list of potential utterances (eg embed -> top k -> choose from top k)
Modern stt is baking semantic EOT into the prediction. It makes sense. Conversation has many potential ends, and you have to deal with misprediction of current words that may signal completion, but mel spectrograms contain a ton of information as well.
In your version, we’d have to deal with partial transcriptions and a system that is not natively audio aware. It would also be much more expensive.
Pipecat already does this. It does it for end of turn detection and it's pretty damn good. It's just that it calls the same LLM I have to do it, potentially expensive and prone to latency and not calibrated.
I think because pretty much every step a business takes is an allocation decision: where to put money, people, time. But it will really pay off only when the models of companies themselves become probabilistic, and the decisions are explainable, at least to some degree.
> Unlike full language models, I don't think the primary business of Typesafe will be serving Jev at API pricing; it'll look a lot more like putting Jev at the center of a much more expensive suite of software.
Precisely this. Should be top comment.
Also, the model moat is understated as training data for these purposes also accrues to the winner, which due to the first mover advantage as well as the distribution advantage you speak of, is typesafe. In contrast to relatively open coding data. Openai anthropic also have that, but like you say its a different business.
At least in theory (TypeSafe has been pretty close-lipped about the details so this might just be hot air, and I think the evidence is a bit spotty) this is false, since they use a different reinforcement training method.
If the word “calibration” in probability doesn’t mean anything to you, the difference isn’t very apparent, but that doesn’t mean it doesn’t exist.
It's calibrated where it's calibrated and not where its not. Their "confidence" score is rather bunk, so it's hard to say. As with other LLMs, you have to eval it for your setting, and if you find it calibrated, then great.
Interestingly enough, at lunch this week it turns out that one of my buddies work colleagues has actually developed a relatively well calibrated version of this, so it's not impossible.
It's probably much harder for arbitrary data though (like most things).
I think you aptly describe the benefits this approach can offer.
What I haven’t seen answered is why it isn’t reproducible within very short amount of time and low effort.
Sure, the large labs are the lazy incumbents at this stage. But any other startups would be able to copy the experience. What’s the barrier of entry that I am missing?
Yes, the best way to think of a general classifier like this is like a smart switch statement. Essentially a "JEV" like thing becomes a sort of programming primitive. Once you see it, it's hard to not get excited.
But even if others surpass them and make better solutions, the fact that nobody was able to see it before typesafe is a testament of what they might be able to come up with next.
I sound like a fanboy but I swear I 'm unaffiliated with typesafe. I was building my own version of this way before they announced JEV (mine was ALE and it was mentioned here on HN for a bit), in use for VR gaming (so one can give commands to NPCs with voice and supports multiple commands in sequence in a single pass), but I missed the "killer usecase" of being a new primitive, like everyone else.
TLDR: There's a lot of value in thinking ahead and seeing the future. The clones are nice and exciting but they give me "I could have built this first, yes but you didn't" vibes. I hope they manage to keep it up and push the space forward again
> [...]the fact that nobody was able to see it before typesafe is a testament of what they might be able to come up with next.
Counterpoint: there are a lot of one-hit wonders, and they vastly outnumber the idea-factory people. This is not to minimize those people, a single idea can be very successful (see Zuckerberg), but it doesn't mean your subsequent ideas will also be great (see Zuckerberg)
You’re correct, it’s a great solution for a very narrow set of high volume classification needs that require very low latency. But that’s it.
I keep evaluating Jev for my product because of the hype, but the reality is that for my tasks, Luna is more accurate, only a little more expensive, and the latency doesn’t matter. I’d rather spend the extra $50 / month or whatever than have to shoehorn in another API and provider, and also lose the ability to change reasoning level and get reasoning summaries for eval purposes.
Did you look into self-hosted open source decision models because they are quite capable also much faster and much easier to integrate and you don't need to pay anything.
I keep hearing Jev is better for accuracy, and still in my use cases Luna beats it. Latency doesn’t matter to me. The cost for additional Luna tokens is negligible for my use case. Idk, I just can’t figure out where to use it instead of an LLM.
LLMs have higher accuracy and are more versatile, but that comes with a cost (slower and more expensive to run). On the other end you can have your own mini classifier something built on top of ModernBert or DIET or even a hash classifier (these are narrow).
Models like JEV fall in the golden middle. Somewhat cheap, somewhat fast, somewhat general.
Many people have replied saying that narrow classifiers (eg bert) are better, and they are better in many ways (essentially free and faster). But... the real world is messy, in my experience it's actually very hard to make a really good classifier that will work across a specific niche domain especially when there is very fuzzy input. And the most interesting real world scenarios are nuanced and overtime there will be edge cases discovered where it fails and that will require retraining and re-evaluation etc etc. It can becomes a full project on its own that eventually collapses into a game of whackamole (improved in some direction but regressed elsewhere).
Things like JEV (or similar models) provide ease of mind, just off load the complexity to it and move on to the next challenge type of thing.
Jev does not have the properties you’re talking about. It’s not very good at general knowledge, it frequently approaches chance level classification on business domains, and there is so much room for someone to do better.
Doing better, I suspect, will be throwing a bigger model underneath and doing something diffusion like (we’ll allocate a whole block to compute rather than just the constrained space) that doesn’t compromise on latency.
>> TLDR: There's a lot of value in thinking ahead and seeing the future. The clones are nice and exciting but they give me "I could have built this first, yes but you didn't" vibes. I hope they manage to keep it up and push the space forward again
This all sounds intelligent and likely, and yet we can come up with countless counter examples where the first mover is not the big winner, and nobody cares about who did it first. There is typically way more value in nailing the execution of a big idea someone else came up with, rather than "seeing the future".
If JEV is here to stay, we can have another split along the "one LLM rule them all" regime
name it worker or implementator or something
- mostly for agent with with spec (not directly for human, as we tend be handwave with vague intent)
- follows the instruction verbatim
- lives in sandbox by default
- basically sol6; but WITHOUT magical situational awareness, lets work in group BS
Yeah, agreed. A model or primitive on its own has no moat and frankly limited value. The paradigm behind "System One" models on the other hand is potentially huge.
Skimming the post, it seems to argue for reconstructing the very rigidity that LLMs let us escape, and that very aspect of LLMs is what made them useful and explode in popularity so much.
One may consider it "rigidity" but many application backends (including LLM powered ones) never needed the full unconstrained capabilities of an LLM. What people needed was the zero-shot capabilities over tasks as opposed to have to train a classifier, span model, etc for every task.
Constraints can be an advantage in many systems. Consider typed and untyped programming languages. Many advantages with typed languages over the latter in terms of development and efficiency.
As I recall, they picked Deno over Bun during the whole vibecoding rewrite fiasco from six months ago, and there wasn't a ton of explanation given beyond the general vibe of "ai bad". They might reevaluate that in light of things being generally fine and losing their precious handcoded runtime choice.
> there wasn't a ton of explanation given beyond the general vibe of "ai bad"
I'm very bullish on AI, but vibe-porting the piece of software you're the main maintainer over a few weeks without letting anyone in the community know and pushing that as a fait accompli to both your userbase and your open source community is definitely the kind of behavior that makes you not trustworthy enough to depend on.
as a counterpoint, your fears are hypothetical so far - nobody on earth cares as much about bun as the bun team and if it was good enough for them it is probably good enough for 99% of users
i think someone who has built their infra on top of bun and have bills to pay care much more about bun stability than the team who get paid by Anthropic to vibecode an entire rewrite to another lang
I totally thought the other guy had you in the argument, and your retort blew my mind. An excellent point, people really need to research their runtimes before placing the entirety of your infra on them.
Then again, I suppose that’s partially the reasoning for the existence of support contracts.
I’m just going to add this here, but: was this whole thing just an aquihire? The dude is a badass, so I can see buying deno just to get him… skillset matches the areas cloadflare has roadmapped out, but still needs to shore up (a lot)
The vibecoded rewrite led to the deprecation of support for Bun but I believe Deno was already the default. Also Bun had the note “No permission restrictions available. Scripts have full file system and network access.”
If I remember correctly, Deno was always priority number 1 due to having a permissions system. Bun's removal due to the rewrite had people complaining using Deno as the focus as Deno's development shifted towards being LLM heavy. Node now supporting permissions would likely be the priority now.
There was a ton of criticism beyond "ai bad". A complete rewrite, in a different language, is a brand new project. It doesn't have any where near the hardening as the previous version thousands of deployments. If they had done that exact same thing but without AI people still would have been upset.
> As I recall, they picked Deno over Bun during the whole vibecoding rewrite fiasco from six months ago, and there wasn't a ton of explanation given beyond the general vibe of "ai bad".
Your comment is baffling. Either you lost track of the story or you've opted to post a very simplistic take on the whole Bun fiasco. Bun's ill-advised rewrite had zero technical grounds and the radical drop in release cadence in spite of all the AI backing suggests the project's foundation lays on shaky ground.
I don't think that's a good interpretation of what happened.
Bun's release cadence dropped because they had just ported to a new language and they weren't about to release that to their community without letting it settle in for a couple of months first - which they did, by running it in Claude Code on millions of computers.
For context, the Rust rewrite started May 4th, landed on main May 14th, the last remaining Zig code was deleted June 24th, and the first stable Rust release was cut on August 19th. Prior to that people had to test the daily unstable releases.
I think it takes quite severe motivated reasoning to look at a project that ported from one language to another and shipped to millions of end users without incident (Claude Code users didn't even realize they were running the new version) and conclude it was a "fiasco".
The fiasco was that a lot of people were unhappy about it. I'd be interested in seeing a Venn diagram of people who complained about the Bun rewrite and people who don't like AI-assisted programming. I suspect the overlap is heavy.
> Development stalled? Here are the stats on commits to Bun in the past six months
You know you're putting up bullshit metrics, not much different than counting lines of code touched by commits.
If you were forthcoming and showed relevant metrics such as release cadence (i.e., what projects declare to be relevant units of work to be delivered to their userbase) you'd be faced with the realization that since march (bun v1.13.14) Bun only managed to put together 4 meager releases:
* Bun v1.14.0 - Aug 20 (first release in 5 whopping months?)
* Bun v1.14.1 - Sep 4
* Bun v1.14.2 - Sep 5 (emergency regression fix)
* Bun v1.14.3 - Oct 10 (just a few hours ago)
This means in the last half year, following the rewrite the project barely managed to put together 3 patch releases. In semver terms this means no feature work, mind you.
Considering one of those releases was a full rewrite to a new language yes, I do consider that productive.
There were a ton of new features in 1.14.0 - most of the release notes were about new features, the Rust thing was a side note: https://bun.com/blog/bun-v1.4
> It adds Bun.Image, Bun.WebView, Bun.markdown, Bun.cron(), Bun.Terminal, bun run --parallel, bun test --parallel, bun audit fix, bun dedupe, and bun prune. And it rewrites Bun from Zig to Rust.
> Considering one of those releases was a full rewrite (...)
I'm not sure you're you're being serious. It feels you're desperately trying to be disingenuous to avoid addressing the elephant in the room.
Listen: this discussion is on how Bun's release cycle stalled due to the rewrite and after the rewrite.
I repeat: Bun 1.4 was released way back in August. Since the rewrite the project barely managed to release two patch releases,and one of them is a hotfix mere hours after the first patch release. Being able to put together at best two patch releases in two months, specially after a major rewrite that is renowned for failing to follow idiomatic Rust practices and in spite of a bottomless AI budget and a militant attitude towards AI, is a colossal red flag. There is no way around it.
Anthropic acquired Bun in December 2025, so there are two questions here: how much did velocity of feature releases change after the acquisition, and how much did they change after the port to Rust?
One potentially useful
way to measure that is to use Bun's TypeScript API definitions as a proxy for "features".
We can look at when those landed on main (which should make them available in a canary release) and when they landed on an actual release.
Actually, to be fair, the real fiasco was project governance.
If I had built my own project on Bun specifically because it used Zed, and had my own custom Zed extensions, I would be justifiably angry that the project ditched Zed without any community input at all.
If that's your complaint than I think it's credible.
Its sad watching them try to keep up when they want nothing more than to crumble under the weight of their own bureaucracy, and at this point I'm not sure there's any hope for them.
reply