A few days ago I was just thinking that Google never talked about their diffusion text generation model after demoing it at I/O a year ago. The rumor is that it was too expensive to run, but with the provided chart using the same 1x H100 hardware and comparing DiffusionGemma to regular Gemma, that shouldn't be the case. I'm curious what the downside for this speed is here aside from being slightly weaker than Gemma.
> I'm curious what the downside for this speed is here
"DiffusionGemma's speedup is designed for local and low-concurrency inference. In high-QPS cloud serving, autoregressive models can be deployed to saturate compute efficiently, so DiffusionGemma's parallel decoding offers diminishing returns and can result in higher serving costs"
Well with a standard autoregressive model you can generate for example 256 tokens at once if you have 256 users, with this approach you can generate 256 tokens for a single user but you need several forward steps.
So the diffusion process takes more GFLOPs, if you have enough users you can already balance memory and compute.
Even if the ReAct paper was published in 2019, I don't think GPT-2 was robust enough to actually work with a tool-calling approach even when finetuned.
For regular coding, GPT-2 was effectively useless because it was only trained from links posted on Reddit.
The dangerous use cases back in 2019 were spam and phishing and GPT-2 1.5B was nowhere near good enough to do homework assignments. No one envisioned how LLMs would develop.
Back in 2019, it was more fair to have caution around the larger GPT-2 models since robust text generation (by 2019 standards) was a complete unknown. For something like Mythos in 2026, where now the social implications of better LLMs are more understood, it's more fair to call it (EDIT: specifically, the declaration of its danger) a marketing gimmick.
This is a natural follow up question -- what kind of an escalation or message should frontier labs/companies publish to be seen as genuine and not marketing gimmick?
I'd say it's almost impossible at that point. Specifically, Altman said so many lies in the past that people stopped believing anything he says.
I think the core of this distrust is the fact that these companies positioned themselves against humanity from the start by saying people will lose most jobs etc. Not only it didn't happen, but many people feel several aspects of their lives got worse because of LLMs, in spite of obvious advantages. So the distrust and reluctance are real.
It's fine for the labs to publish model safety cards and stagger releases/limit it to a narrow test group as they are already doing, but saying they're doing it "because the models could be dangerous" comes off as unnecessary as best.
One of the main purposes of model cards, from the beginning, has been to outline the ways that a model could be harmful or dangerous, and mitigations that can be or have been taken to reduce those risks. How do you expect labs to publish model cards without talking about this rationale?
What? So it’s fine for them to be concerned about the safety, try to measure it, publish results about it, start with a cautious phased release approach, basically all the things they’re doing.
But if they say why they’re doing all those things, it’s too much? What?
I dont think they can, their rhetoric is so silly these days that if they did accidentally release something dangerous no one would believe them. The Boy who cried Rokos Basilisk I guess.
These chicken littles have lied far too many times in far too extreme ways in their desperate attempts to obtain a monopoly by regulation.
The remaining 'worthwhile message' would be that they have deleted their models and are dissolving the company, because they believe the risk was too great and was worth losing the revenue and risk being civically prosecuted by their investors, and will take the chance that they'll be able to convince a court/jury that they acted properly.
In other words, putting their own skin on the line for the veracity of their claim-- rather than everyone elses.
So the people who tried to kick him out of the company now have a bunch of vague statements like there being "a pattern of behavior related to his honesty and candor". No other details. What a crushing indictment.
Has it been released to the public yet? Genuine question. Because if you didn't try it yourself, you have to rely on others' reports. And different people who tried it on different projects got different results, leading to different conclusions.
I'm happy to have a discussion with you if you bring any argument.
Before GPT what would have been your choise of architecture, setup, alogorithm if someone comes to you and says "write a tool/system which can generate code" "what do you mean generate code? How do i control it?" "by writing what you want in natural language" "puh 50 years of development, 100 billion, top tier team of linguists and software engineers perhaps?"
Ask StackOverflow if they think it didn't change anything for them.
Programming is the reification of decision-making processes. If you don't understand the decision-making process that you want, you get a different one, which at best approximates the one you want but couldn't articulate.
If you do this with COBOL or Python, at least you get consistent operation and errors when you're wrong. If you do this with any LLM, consistency is dropped in favor of obsequiousness.
The base problem is that people aren't equipped naturally to think about all the details of their problems.
It’s mind-boggling that anyone could deny this in mid-2026. Virtually every software engineer I know is no longer writing the majority of their code. Many are not writing any code, myself included. And I’m a staff engineer with 20 yoe, formerly at big tech, and now building a (profitable) SaaS of my own. The way I work is wildly different from a year ago.
You can scaffold out a simple app pretty easily. Anything large or complex things break down. If you don’t know what you’re doing you end up leaking secrets like the dozens of examples we’ve seen so far.
On one side, it means that a certain amount of business will just use it even if you think its not safe/good enough and they will throw out people and will still succeed.
And on the other side: yes because they will also use LLM review or other tooling and will be fine whatever the 'security llm agent' tells them.
Before gen code killed the freelance business model, there were hoards of people on Upwork/Fiverr willing to fuck other freelancers over and underpay themselves to make whatever barely-working slop you wanted.
Hell, before managers got the idea of AI layoffs, they had been off-shoring to low-quality code sweatshops for years. That was supposed to kill software engineering in the States 20 years ago. And it was just as frustrating (if not moreso) to get them to actually fulfill the project requirements.
There is almost no maintenance work for bespoke apps apart from infrequent updates to keep OS and hardware compatibility as the environment slowly changes.
Keep in mind, these are not products in the endless feature treadmill promoted by scrum.
For starters it makes you able to bypass having to go on Reddit to find incomplete trace of solution to some niche problem and acts as a sophisticated (but sometimes wrong) search engine. This already is worth every penny and improved my mental health immensely.
Style is relevant to how humans communicate and it's not always about the message, and it can sometimes work against it. AP Style is an editorial standard for a reason.
IF I WROTE AN ENTIRE BLOG POST IN ALL CAPS ABOUT HOW AI IS LITERALLY SATAN PEOPLE WOULD JUST THINK I AM A CRAZY PERSON
Content with a heavy smell of being AI generated already gets flagkilled. Vibecoded repos on GitHub don't strictly fall into that bucket since they can still provide functional utility.
Do you have a specific example of something that should not be allowed but did not get flagkilled?
Separately I've gotten down voted for pointing out something is AI generated, but if it's explicitly policy that it isn't allowed, it seems like that should be written down.
For pedantry's sake, they were saying "AI = Apple Intelligence" last year as well, so it's not like they just pulled it out of their butts now that popular opinion has turned against AI.
reply