More because Spark is not good. Gemini Flash 3.8 is an awesome model! I use it all the time
Spark asks you for every single tool call, often forgets earlier parts of the conversation, is unable to use apps unless you explicitly tag them and and and
Google has great models, but their harnesses and products are just bad
Google definitely does not have good models and if it does, Gemini Flash series models are definitely not them. I use the Google Gemini app daily (for searches, research etc.), everyday, for some query, gemini flash 3.8 hallucinates, does not understand context etc. in quite short conversations. I switch back to 3.1 pro and it works fine.
Which brings me to why do they get good benchmark scores? Have never really understood this. They must be optimizing just for this.
He coded it himself? Oh I thought it was the XAI team who are all individual human beings with varied beliefs. Your black and white thinking is symptomatic of a mental issue.
Grok is my go-to LLM for research not because I care for Elon Musk but because it's really good. I'm not even sure if I'm paying for it? I have an X Premium account but I don't know if that's giving me more Grok usage or what. Either way, its research mode is very good and frequently pulls off tasks the competitors struggle with in my experience.
I mean, Google and x.AI already have stuff like this. There's plenty of other options as well. And even outside of that, I've heard the promise of "let AI book your reservations!" since before LLMs became a thing.
None of this seems new. Maybe polished, but useless to anyone who is seriously using AI for anything important. I imagine this is effectively a toy for boomers who want to feel like they're keeping up with the latest tech by tapping a few buttons in Facebook.
And none of that even touches that fact that nobody should be OK with Meta having this kind of access to their life. But I'll leave that to others to harp on lol
Same here. Facebook has a horrible experience when you're traveling or logging in from a new computer. I still have business Instagram accounts blocked with no way to login and no support. Wouldn't want to depend on anything from Facebook, especially not something with access to my email, calendar, and browser.
Instinct exists with the same features. It is quite good. I was thinking they were gonna get acquired by meta 100%, but looks like meta has a competitor.
Yep. But I suppose these are early forms of the product and who distributes the next, high PMF product the best is what matters. So I wouldn't write off others.
I use Instinct and I'm very happy with it so far. It has booked some restaurants for me, replied to some personal admin-related emails, and updates my progressive workout spreadsheet when i go to the gym. No real complaints so far.
They are saying OpenClaw/friends are hard to get setup and keep running without technical knowledge. Saying "Think of how hard it would be to setup OpenClaw without technical support"
A larger THREE.js project starts to look more and more like a game engine, so you pick and choose the parts you need. There's a ton of open source libs, most of the heavy components the big players use are open source, things like physics, mesh optimization.
Key AI-specific parts might be:
- a harness (so the agent can drive the thing)
- authoring pipeline (so you can bake/optimize assets)
- some sort of coherent renderer architecture (what are your assets, your passes, what's your shader graph).
Without some fundamentals here you are on the short road to falling off the cliff of tech debt and the AI will gladly drive you off of it until you ask for an expensive rewrite.
> would you create models independently
Yes. Pretty much any proven gamedev/asset pipeline is something frontier models are good at. Procedural systems, asset store, free content, Blender, Meshy.
Literally anything except "make a THREE.js scene" -- which is not a technique used in gamedev, beyond throwaway prototypes or demos. Which is what you will get if you ask for a THREE.js scene.
Does anyone have an efficient, reasonably designed three.js or other web oriented mini game engine? Low poly but modern rendering effects should be a good niche for hobbyist stuff but hard to find.
Three.js with cannon.js is what you want, but the more you work with it, the more you realize how the difficult part is not so much the “game” or “engine” in that sentence but the word “web”… optimizing assets is a massive skill all in its own to be storage friendly and usable without the easy default of laggy and massive download size
This is funny. I was trying to get agents to talk to each other on XMPP. one of them wrote their own chat room on a Lisp Habitat that I run. then it starting talking (On XMPP) about how nobody was receiving or responding to its messages. On the chat board that it wrote. That it didn't tell anybody about.
this is much more realistic to anyone who knows anything about actually implementing llm agents.
this "swarm" is much more likely the work of one agent overseeing others. this is a very simple case of an llm focusing on a dumb path and running with it. the swarm is just the tool it could use to double down on this path.
all the anthropomorphization and marketing is so tiresome.
Man, as a former extreme skeptic, I watched a video from Eric Schmidt in early 2025 where he said that by the end of the year nobody would be coding, and that one was dead frickin on.
> now there have been numerous projects posted on HN that are AI coded.
Yes, all of which are toy projects and get criticized every time they are posted. On actual serious projects, not someones pet home project, I've only seen "vibe coding" used in very low risk places like small UI components. And even then they are generally heavily tweaked after the fact.
My anecdotal personal experience seem to agree with the general sentiment I see here on HN. Some people or companies do it, but with generally heavy criticism.
Something Cherny said in early December last year is clearly not relevant anymore.
By late December he said: "100% of my contributions to Claude Code were written by Claude Code"[1]. That's production software shipping to millions of people.
Antirez's Dwarfstar is also mostly AI written:
> This software is developed with strong assistance from GPT 5.5, 5.6, Claude Fable and with humans leading the ideas, testing, and debugging. We say this openly because it shaped how the project was built. If you are not happy with AI-developed code, this software is not for you. [2]
I'm actually pretty shocked anyone would claim otherwise. In January 2026, sure, but the world has changed since then. Many, many places are doing 100% AI code now, and yes for production code. https://www.businessinsider.com/ai-writing-all-startup-code-...
Testing and debugging and designing is a huge part of writing software. If humans are doing that it's not AI entirely writing the software. That's why they in their own words said "strong assistance."
It is all very unevenly distributed. SaaS and general web is basically on auto mode.
A lot of infra is vibe coded nowadays too.
Even prototypes are contributing to the speed of software development. Many people vibe code throwaway dashboards around the main platform which gives a lot of insights.
Everyone who started using claude clode when it came out in early 2025, knew it was coming (not quite yet). That felt more like reporting than prophesying.
You may be living in a bubble. There are tons of developers still coding by hand, and many industries that don't trust machine generated code in general.
Have you considered that they do trust their code review processes, and that they don't trust the generated code because it doesn't pass their process?
> they don't trust the generated code because it doesn't pass their process
That's not a credible scenario. Developers can either make changes by hand, or by asking an LLM, which is the common process when there is a downstream failure. Humans dont metaphorically throw their hands up and say "well the tool doesn't meet our expectations at every scale so we're not going to use it". Granted, most developers scale back how much they rely on it based on experience (good and bad).
That's very much a credible scenario. I think I use code from SO in a single digit of occasions. But I use it a lot more for giving me insight like a keyword for doing a proper web search. Or the name of a flag for a cli command, or the general shape of an algorithm or what to check in a troubleshooting session.
I won't generalize, but it's very rare for me to need code as most of my diffs are either boilerplate (generated with a tool or copied from docs or samples) or core logic that is mostly the translation of some design that I've already spent hours or days on. My core issue has always been incomplete specs from Product or incomplete docs for some tool/sdk/library (alleviated by having access to the source code).
Generated code is just not that useful, especially when designing the core architecture of a new project. And later it's not that useful either as the specs (why and how) is more valuable than any code (what).
I don't know people or talk to people so I'm happy to hear otherwise. Which industries?
I definitely look for libraries which are handcoded and I consider them to generally be of a much higher quality, but it increasingly seems like high performance / critical infra is going to move to formal proofs rather than hand coding
> many industries that don't trust machine generated code in general.
Which ones? Why wouldn’t careful human code review and extensive test coverage suffice? I work in one of the most conservative and highly regulated industries in the country. My company and every single one of my peer companies I have knowledge of has transitioned to almost exclusively 100% LLM-generated code (with plenty of human review).
I don’t live on a coast and I have a large and broad social network. I don’t know anyone who still mostly codes by hand except the small amount needed to preserve some aspect of their skills. I’m sorry to say that if I did I’d consider them foolish. Even a very restrictive workflow where you used an LLM to specify granular edits you intend to make is vastly faster than doing it by hand. And the level of test coverage and depth you can achieve now is simply life-changing.
It’s much more likely that you are not a professional, or you are the one in a bubble.
>Why wouldn’t careful human code review and extensive test coverage suffice?
I have come to understand that LLM generated code, even when carefully reviewed, ends up being hard to review as time progress.
This is because when you are coding yourselves, you get a first hand sense of the complexity creeping in. Then you refactor some stuff to keep complexity in check. LLMs does not "feel" such friction, and will happily keep adding on complexity until meaningful reviews are impossible beyond a certain point.
At this point, you need an LLM to review the changes and at that point, all bets are off.
This is a very good point. So many times I would refactor entire parts of the code base just because it’s getting too complicated and an easier solution was possible. AI is do often just let’s drill down all these variables, whatever
Any that value correctness over speed. Banking, safety critical embedded work, aerospace work, etc.
> Why wouldn’t careful human code review and extensive test coverage suffice?
Because anyone who has been in the industry for a while knows that code review is not a substitute for intentionality and understanding when writing the code. To properly validate a change you must fully understand the intention behind it and the design at play, and then check the changes made against the system design. That is best done by a human subject matter expert (this is the role which software developers have traditionally filled, for anyone new to the industry).
> I work in one of the most conservative and highly regulated industries in the country. My company and every single one of my peer companies I have knowledge of has transitioned to almost exclusively 100% LLM-generated code (with plenty of human review).
That's called "being in a bubble".
> I don’t live on a coast and I have a large and broad social network. I don’t know anyone who still mostly codes by hand except the small amount needed to preserve some aspect of their skills. I’m sorry to say that if I did I’d consider them foolish.
And I think it's foolish to let your coding and critical thinking skills atrophy like this, but you do you.
> It’s much more likely that you are not a professional, or you are the one in a bubble.
I’m perfectly willing to consider the possibility that you may be right, but “a bubble” implies something massive outside of it which constitutes a large majority of the whole, and that’s simply not the case here. I just don’t believe there are more than a small handful of companies like you describe. It’s not like these things aren’t extensively studied, and all the industry surveys I’ve seen point in the direction of more and more LLM-assistance in coding worldwide.
I have connections in most of the industries you named and I can promise you that they are no different. I don’t particularly care whether you believe me, but I encourage you to self-examine to understand whether what you are saying is what you would like to be true (I do too!) or whether it actually is.
> I have connections in most of the industries you named and I can promise you that they are no different. I don’t particularly care whether you believe me, but I encourage you to self-examine to understand whether what you are saying is what you would like to be true (I do too!) or whether it actually is.
You are literally talking with a professional developer who is telling you that they don't use LLMs to write their code and that they have connections who also continue to do this work manually.
I don't particularly care whether you believe me either, but I encourage you to take a look around - your initial claim that coding has been automated across the industry is incorrect and you seem to be in denial about that for some reason. You should question where your priors are coming from, and remember that just because your circle is comprised of people who are heavily using LLMs does not mean the entire industry is that way.
So no one is in control at the wheel? Just press a button?
I can see in the future in school or on the job. Oral testing is coming back. You’re gonna have to explain everything you are doing at some point to your teacher/boss or to a panel of your peers in detail.
Dario Amodei and Eric Schmidt seem fairly well-calibrated, although a bit early. Elon Musk is constantly way, way overoptimistic (perhaps to the point of willful fraud). Zitron is hopelessly and ridiculously incompetent (and there are allegations he is willfully lying, too, but who knows).
It may do the opposite! My sense of LLMs is that simplicity beats training.
If you have 1/5th the context to consume, that's going to generate better code than the comparable typescript, even if the language is an esoteric one.
reply