I think it is ever more important to realize who is releasing models rather than what the models do and how they compare.

Because models iterate at breakneck speed, looking at today's benchmarks is only useful for someone using the models today. Whereas if one builds a product on top of it, or commits to one for a project or team, the company or organization behind it, is far more important. Will they exist in a few months? Do they need a business-model? Are they subsidizing usage with venture capital and how long can they keep this up?

Is reflection a company? University lab? NGO?

> The startup was launched in March 2024 by Misha Laskin, who led reward modeling for DeepMind’s Gemini project, and Ioannis Antonoglou, who co-created AlphaGo, the AI system that famously beat the world champion in the board game Go in 2016.

with the obligatory:

> Investors in Reflection AI’s latest round include Nvidia, Disruptive, DST, 1789, B Capital, Lightspeed, GIC, Eric Yuan, Eric Schmidt, Citi, Sequoia, CRV, and others.

If you get access to a VM, it's not a "security vulnerability" if you then have access to that VM. This was the whole point, the product.

It's almost like returning a car after you bought it with the reason "When I open the door with my key, the door is open and anyone can get in".

Some models are reproducible, in that the same prompt will generate the same output. Say that we could wire up such a model to generate some code.

In that case, we could create a prompt that generates, say, an entire codebase, or a large piece of text. The prompt (or really, the tokens) would then be the compressed version of the codebase or the text.

I am not talking about an "AI agent", but really a model that we call in a reproducible manner. Preferably one call, with one prompt. An agent could just run `git clone` to "decompress" a codebase, which conflates the idea of compression. If that were compression, then the "compressed version of the git kernel" would be a single line of text: `git clone https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/lin...`. I am really talking about having an LLM re-generate text based on a prompt.

Does that make sense? I can imagine that this is highly impractical and inefficient. But would this count as "compression" at all?

A large language model itself (the network) give you the probabilities for the next token given some prefix of tokens so far. You can use arithmetic coding to go from these probabilities to a deterministic compression / decompression algorithm.

When you use an LLM to generate text, you sample from that probability distribution. You can use a true random sample. Or you can make it trivially deterministic by using a seeded pseudo-random-number-generator or you just pick the highest probability each time. But that's all a red herring; really, what you want is arithmetic coding.

https://en.wikipedia.org/wiki/Arithmetic_coding

A chat?

>I am not talking about an "AI agent", but really a model that we call in a reproducible manner.

An LLM is just as deterministic as any other computer program. For identical inputs (which includes the PRNG seed) it produces identical outputs.

>compressed version of the git kernel

The git kernel, got it.

>But would this count as "compression" at all?

Yes. The decompressor is several tens of gigabytes though.

This is not really true in practice because of multi-threading and out-of-order execution. Mathematically equivalent orderings of operations are not equivalent when dealing with floating point values, so most practical LLM implementations end up being non-deterministic.

If you wanted to and had enough engineering effort to spare, you could run an LLM deterministically at relatively small impacts to performance.

One approach is to make sure you run things in the same order. Another is to change your operations so that more of them become associative or even commutative.

See eg the paper 'A Lattice-Based Approach to Deterministic Parallelism' for some interesting ideas on the latter.

https://imalogic.com/blog/2024/06/03/image-compression-decom...

Like when you click the Calculator button on your android, it wouldn't actually exist yet, your click actually prompts it into existence. But naively that has problems because you don't want a different UI every time. There's something to your idea.

https://youtu.be/7NfyZhV1dKM?is=YOUXCHuFUiPdlD0p

The purpose of a democratic government is to reflect the will of the people, not the other way around.

Fighting misinformation has to happen through establishing and maintaining trust, and providing trusted counters to misinformation. Establishing trust requires alignment with the people you want to trust you.

Interference (as opposed to participation) in the free flow of information and the process by which truth is determined is a great evil.

I think it's actually a feature when people can recognize it as AI.

Why would someone want it to be unrecognizable as AI? What is the reason people want to hide AI usage ? (other than when "cheating" in studies or examns or such, obv)

Both of these communicate that I will probably not get a good value for my money when buying your product or service.

This is where I stop following. Businesses should focus on where they differentiate, and I really don't care how they generate signage if I can read it. I don't get anything by them paying Banksy to make their menu instead of having Gemini generate it. And if there is any connection to getting good value, it's the inverse: I'm more likely to get good value when I'm not paying for fancy advertising. See: every drugstore that puts generics alongside brand names for cheaper.

The worst generative AI posters and restaurant menus I'm seeing from newer small businesses do not instill the same feeling that I'm getting generic-like value. To keep with the drug analogy, they feel like the combo pills you find somewhere in the back of a gas station which combine 10+ random supplements to help with 20+ different symptoms. Especially for new restaurants where real reviews are limited, menus with terrible AI generated food drawings really give me the same ick - that they might not be providing any good value at all.

However, i do admit that part of this is also that i live in an area where there is tons of competition for most cuisines I'm interested in, so I can use menu quality as a differientiating factor in the first place. If enough places around me were using sloppy AI menus i would probably ignore it.

I'm saying I don't particularly value businesses that focus on things that are low impact for me. I'd rather my barista focus on sourcing good beans and making good coffee than paying someone to do a fancy custom menu for them. We see this all the time when we visit new areas. It's considered naive to follow whatever looks most glitzy -- every seasoned traveler will tell stories about all the tourist traps, and how the real quality is the hole-in-the-wall place all the locals gather.

What I'm saying is we shouldn't conflate quality with marketing and advertising, which is what posters are. You seem dead set on doing so anyway, and I'm not sure why.

If I send a single email "subject: poster" "make me a poster for our event, details below, we need 200 printed" that's also low effort. If I do this on fiverr, it's also "unwillingness or inability to spend resources"

> your product or brand.

I think you are conflating a wide audience with a narrow idea. We aren't talking about just the large coorporations, or the popular rockband. There are local bookclubs, a carbootsale, a group of hobby-foo-bars showing their foo-bars at the local pub next month. A concert by "amateurs". Someone looking for an appartment. A cat that got lost.

Do all of these need to show high-effort and/or willingness to spend resources?

AI-generated visuals aren't the first to go through this. I'd say the early-to-mid 90's "desktop publishing" went through this too - as soon as everyone could make ads and flyers cheaply with Word, everyone did.

If you can tell that it's AI, that's probably because it is low effort. Higher-effort attempts pass more easily.

(Of course, that also requires an understanding that there are ways to apply effort to making a short request; and believing that it would make a difference; and noticing for yourself that the first output is bad; and trying. Quite a lot of people seem to lack at least one of these.)

It won't die off until AI stops being used with low effort. Which it is the majority the time used for.

If a business markets itself with these AIgen posters, you get wildly different responses from different groups.

boomers that spend all their time on Facebook will think “this is such a nice pretty poster” and be oblivious to the fact it is even ai. Even if they knew it was ai they don’t care.

Young people will think it is low class/trashy. Some hate it because it’s ai, others hate it because the poster is ugly/busy/confusing.

Among many other things, it's about "am I getting scammed". So my mind needs to develop called ID in some sense. Not a perfect indicator, but enough to protect me. That can open me up to criticism as being overly cynical. And my response is: yes, that's what a low trust society does to people. That's the point.

It's perfectly possible. Another aspect of humanity pointed out to me from others is that we are surprisingly sensitive not to the current "quality of life", but to the rate of change in that QoL. So we don't necessarily need to fix all of humanities woes to restore trust, we need to get on a trajectory where people believe a better future is coming. That optimism and morale makes all the difference in building momentum towards restoring trust.

But whether we should ultimately comes down to your own personal paradigm. Do you feel you succeed when you reach your goals and lifestyle you desire, or do you get more satisfaction helping others succeed and survive where you can? Would you rather feel more secure wandering around your neighborhood, or are you fine paying for extra protection (directly or indirectly) to get by your day-to-day? Do you prefer a variety of public places to take your family to or would you rather have a nicer dwellings to invite friends over to? There's no correct answer to such questions, but they affect the trust of your community.

>how do we get to — if not high trust — not-low trust society when all is good

Not to sound snide, but: you increase trust by doing actions that make people more trustful. There's no magic to building trust, but it's also not complicated. You simply perform genuine, intentional actions made primarily to help others. I won't say altruistic, but the more altruistic the better the effect.

To break it down a bit more, I'd say there's 3 main layers of trust to build: local, institutional, and societal. Unless you have outsized influence already and a passion to lead the charge, you generally need the lower layer's trust improved before it expands to the next layer.

1. in your local community, with people you interact with in your day-to-day. Simply being more available to meeting up with people and doing small actions to assist helps to build up trust. As well as in what kinds of activities and groups you join to meet such people. It's the layer you most likely have the most power over

2. in the institutional level, to influence structures you directly interact with. The school your kids goes to, the emergency services, the platforms you consume media on, the businesses you go to regularly, your local town initiatives. Having trust that these systems (at the bare minimum) aren't actively trying to make you or your family worse off builds trust that the structures around your immediate community will improve and benefit everyone. Many of these are influenced by local policy, either through coalitions to get such proposals to leadership or getting leaders in who care about such factors. They can't do something like giving people a 4-day workweek, but it can make a park cleaner or create more/better public transportation.

3. Then on the societal level, to influence wider culture well beyond your sphere. This is all the larger level thinking concerning economy, technology, environment, this topic here on the trustworthiness of AI generated posters, etc. that the internet often talks about (which make sense, as a medium well beyond your sphere), so I won't need to expand on this much. This is pretty much exclusively larger policy issues to enact in laws, so influencing wider state/federal law requires larger populism movements.

Especially when you see a lot of talk and no walk from everyone around you — yes, I could also be the agent of change, but life sometimes has other ideas.

Indeed. Even 2 truly altruistic individuals may answer the questions differently simply because one comes from an individualistic society, and one a collectivist one. One may genuinely believe cheap, private self driving cars is a future to strive for while the other believes in efficient and fast public transportation. both can accomplish a good future with the right intentions. And maybe both to some extent does work in tandem. But we know land and budget for roads vs rails is somewhat competitive.

We definitely lack the energy and authority to tackle all issues, so I simply want to try picking one direction and putting effort into that first. All my energy has been spent job searching until very recently, but once I financially stabilize I wanted to find some local coalitions and see what I can do to chip in. I figure such groups would be a good place to start, people who are already walking the walk.

bit of a problematic argument in this context given that the entire usage of generative AI rests on people's unwillingness to do just that

I want the products of it (how many phones with lithium batteries in those protests), but not the side effects.

Which is not to say these are wrong perspectives to hold, but it's important to understand where the root of the complaint is before you try to come up with generic solutions.

If we operate under the assumption that the AI generated posters are bad (which this HN thread is), then making them not look like AI generated posters would be directionally toward making them not bad. That doesn't necessarily mean it's good though, it could be average and fulfill that constraint.

Because a lot of “everyday” folks really strongly dislike AI. I’ve lost count of the number of AI generated Instagram ads I’ve seen where the comments are filled with “AI slop!”.

It’s not a positive signal for anyone who wants to persuade someone of the value of whatever they’re advertising because it’s associated with being cheap and lazy.

> About half of Americans said in the June survey that AI will worsen people’s ability to think creatively and form meaningful relationships with others. Far fewer said AI will make these things better.

https://www.pewresearch.org/short-reads/2026/03/12/key-findi...

“about half” surely objectively qualifies as “a lot”.

> I can also say "a lot of people really like AI-generated arts" and that would be true as well.

Yes but without the corresponding negative. If someone who likes AI generated art sees some human created art they don’t think “gross, why didn’t they use AI to do that?”. So for something like advertising using AI will alienate some potential customers. Not using AI will alienate no one.

If most people aren’t shareholders but instead workers whose jobs are threatened by AI, we should expect this response.

It has nothing to do with artists, it has everything to do with trying to mount a collective response to the apparent threat, poorly expressed as crowds tend to do.

This is a form of the broken window fallacy. Society is not worse off when production requires less labor, any more than it's worse off when someone uses an open source library rather than paying someone to reimplement it.

Society is definitely worse off when people who exchange their labor for capital have no opportunities to do so

Not objectively in every situation. But that does seem to be the case and objective here.

I’m not sure I see it, can you elaborate?

The narrative we're seeing around AI and current automation isn't "we're going to automate one category of jobs". If it were then sure, we could be relatively confident that the excess labor from that industry would be absorbed by other segments

But that's not what the future looks like to me. The future to me looks like "we want to automate a large majority of labor"

So what then? If tons of labour is displaced then the Broken Window Fallacy doesn't really apply, because it is based on the assumption that there will always be something for displaced labour to do.

Call it the Excess Labor Corollary. If there's no way to absorb excess labor in other segments, then what.

Or of course, a few billionaires could hoard the wealth.

There's a Wikipedia page about it.

We could be in a post scarcity economy now, except that generally instead of realising we all have enough of we share, we raise the threshold of what we consider to be enough comfort and stuff.

It definitely could be. With current trajectory as of now, do you think it will be?

>except that generally instead of realising we all have enough of we share, we raise the threshold of what we consider to be enough comfort and stuff.

I see the opposite sentiment. I see previous basic necessities further out or reach then they were last generation. People aren't raising their standards, they are sinking towards poverty.

But then their worries are dismissed because they can get a flatscreen TV for 200 dollars. A TV that is likely feeding your data back to corporate to make thousands off you.

Like, if we all settled for a 1950s British living standard, using current manufacturing technology, we could achieve that while everyone works a little bit less.

But, we'd prefer to keep labouring and have nicer things.

And your theoretical implies we have choice in the matter. I'm sure many would in fact take a smaller home with less amenities over a modern home with cheap appliances. Not that I'd actually be given that choice given my family lineage, but the general living standard of the 50's was quite high. I doubt the act of AC and the lack of knowledge of abestos are the things that made housing skyrocket over the decades.

So here you never have to pay anyone to fix your windows, because they fix themselves, but conversely nobody will give you money to fix their windows! And then you have to pay your self-fixing windows a rent in tokens without an income.

That's a pretty big leap.

Because the reason for any given AI work being "recognizable as AI" is considered inherently a quality issue. For the suspicion to arise, there has to be something that someone considers "wrong" about the work, in order to point at that thing and say "this must be the work of an enemy 「ＭＯＤＥＬ」".

But only for compute tasks. So, practically an alternative language to write compute shaders in.

I believe many of the problems we have, need social and human solutions, especially when the technical solutions are hard or impossible.

Like here, where we can no longer trust images to depict reality and act as proof. While its admirable that people look for technical solutions, the obvious social solution is to admit that images are no longer absolute proof and will become less and less trustworthy¹.

And, by admitting that, change our relation to these artifacts. Sure, that will change journalism, police work, legal systems, etc etc.

But pretending that we can rely on images might allow journalists, police, judges to continue relying on them as if they're authentic, which is a far bigger problem over a longer period.

Social solutions require effort, demand flexibility, take time and are messy. But this is what humans are and do. Not everything has a technical solution. Not every technical solution is the best option.

¹ we already saw this when people claimed "someone must have hacked my iphone and put it there". For decades we've seen this with images that are deliberately taken in a way to spin a story (like the illusion of a large crowd or spacious room through carefull angles or fancy lenses). And I predict we will see this with security footage, "live streams" or even bodycams with "ai enhancement". Just imagine a bodycam or a dashcam that manipulates the output to benefit the owner. "A dashcam that will prove your innocence in assumed traffic violations" or such.

Not that it requires Google to "open up" their play-store, but that they must allow other app-stores to work on the same level. So basically allowing devs and users to move elsewhere.

Apple moved first with making a special 'sideloading' case for apps not under their control, Google is just copying what they did.

No more free sideloading.

Honest question, because AFAIK there's no guarantee or (legal) requirement to support any API. Whether that's fully documented, has SDKs or whether it's something reversed-engineered doesn't matter WRT the support the company owning the API is supposed or required to give.

Or am I wrong there?