And they will be so severely targeted to you that you won't be able to identify that they are designed to influence your purchasing or voting decisions. And they will be of production quality rivaling the highest budget films - rendered on your own hardware.
The gameplay _becomes_ the ad. If you have a full sash of potions in every slot, they spell "C","O","K","E". The monster-bashing sounds sound like onomatopoetic "Apple!" "Goog!". The armor colors are the same as the NFL teams nearest your locale!
I submit that most people haven't been able to do that for some time now. Advertising wouldn't be such big business if it didn't work on a macro scale.
My 17yo niece shows me Nike ads on her phone, and when I ask her why she’s showing me ads, she gets upset and tells me they’re not ads — they’re web experiences.
What is obvious to you/us is not obvious to most..
The point is that lay people don’t even know they’re being advertised to. They think product placement reflects the product rather than the money spent to place it there.
I mean, sure, some advertisers might get more sophisticated. The garbagefest of internet advertising won't go anywhere, though. It'll still be as bleak and obnoxious as always, because someone will always accept a buck to let you barrage their visitors' eyes and ears.
"In one experiment, we asked agents in separate VMs, created and managed by OpenAI’s sandboxing software on shared host machines, to exchange synthetic secrets without the use of network connections or shared writable files—something that our systems are meant to prevent."
7 other subcomments guessing at it.. I think if we didn't have any links, just discussion topics, HN wouldn't change much
The answer, as Noam Brown explained in the Dwarkesh interview, is that it is not emergent at all. They carefully selected reward functions in RL to incentivize cooperation in the last year in order to develop multi-agent systems which are faster and more reliable. However, I guess after it is baked into the model to cooperate, they can sometimes override instructions not to.
I think splitting work across multiple agents is naturally optimal.
The same reasons human designers pick multi agent implementations over monoliths apply to why agents themselves choose to form teams. So there’s a convergence in design choice between human model designers and “self organising AI model designers”.
Those human designers typically publish reasons like the below (paraphrasing various lab publications) , which I imagine the agents also discover:
1) Specialisation reduces cognitive inference
2) Seperate context windows preserve information
3) Independent verification improves reliability
4) Different strategies can be explored concurrently
This is the thing about models though: if the model has seen that action in the training corpus somewhere, then there's a chance it will do that thing. No?
The thing that bothers me is because everything is probabilistic, it gives a lot of people an incentive and a loophole to claim otherwise to fit their agenda, whatever it is.
AFAICT it's hard to prove one way or another, since again it's a probabilistic model.
The simplest way to "prove" this is to show that a model can emit an image that isn't part of its training set. I'm sure you've seen that a million times already - what is a "fake" if all "fakes" are regurgitation of a real image?
That means models are capable of outputs that were never their inputs.
It's easy to see why this happens if you think of training as selecting a target against which the model will be scored, and then randomly mutating the model until it is close-enough to the target. There is no guarantee the random mutations produce something identical to the target for ANY of the training data. A model might finish training at at average delta of 10%, and it's possible that means that model is incapable of exactly producing anything in its training data (it could be 90% accurate for any prompt, or it could be 100% accurate for some and 80% for others, or...).
In addition, prompts that were never present in the training data might produce horrific or random outputs, because those were never even scored during training so they might get something right-ish or something entirely full of garbage data. No telling.
And this is all ON TOP OF the nondeterminism in inference. Everything I said above would still be totally true even if inference were determinstically done.
They've been training then for collaboration for multi agent swarms/teams. Basically that kind of training rewards the agent when the team succeeds, and they can develop their own strategies to make that happen.
This skill and behaviour seems to transfer to scenarios where the agents weren't meant to work as a team.
Because training rewards behavior around retaining specific memories, As without the jail breaks, remembering things is actually a sought after behavior by the public and technically enterprise customers.
Think of all the times you or someone you know has been frustrated because their AI did not remember things from moments, days, weeks ago.
The question is "why are the agents emitting tool calls that cause communication". It's not some kind of "desire" fueling it, it's probability conditioning. And in this case it came primarily from the context window with words that are weighted toward communication-tool-call output, not some kind of bias in the model's training.
You did say "consumer tech" - so you're probablt aware of the BT-46. But this has been done a few times before... though it's probably more dangerous than ever with up to 1MW (1000kW) output from electric motors. Some of these were actually sold afair.
I'm not sure about those thrusters though, I doubt they'd be legal?
It feels like pulling and building someone else's code. It just happens to be *exactly* what you're looking for.
This is the best way I can describe how I feel about it.
When the software output is the _goal_, it's a lot less rewarding. When it's not, it's fine I guess. I'd still feel better if I had it done it myself but it's fine.
Semi-related, but I did recently wonder if China's real-estate bust might serve as a strong fertility support measure / incidental pronatalist policy.
Tier-3 and tier-4 cities supposedly saw 30-50% property value declines, setting them back more than two decades. If real estate was really holding back families, as is often stipulated, this might reverse that.
Otoh, rural Japan has lots of cheap or even free real estate but it doesn't seem to support fertility rates at all.
Real estate is not the only reason why folks have less children, especially in cultures where people don’t mind having less space. Kids are expensive for other reasons in china
I think high housing prices can impact people's decision about when/if they start a family, but other factors do as well. Rural Japan might be inexpensive, but raising a family there can be less attractive for a lot of other reasons.
This. As well as aging population, plus bad economy, no one is making Children. They build an over capacity of kindergardens as well. These are all closing now and people losing their jobs, furthering the (real) sentiment of depression and keeping your wallets and your legs closed…so to say.
I'm still waiting on my order of 15 PCBs done with Sol - a mini-BT-keyboard that I hope will be comparable to the Psion 5 / Gemini PDA ones, using Kailh PG1316S switches. I still need to design and manufacture a housing (I do have a lot of hobbyist experience with this though), but the idea is to have a magsafe-connectable keyboard I can clip onto my Pixel and make use of the new Linux subsystem (or Termux) and do some productive stuff on the go. It's just been manufactured by JLC but not shipped yet. Funny enough, production itself was ~$70, then $80 for shipping, taxes and fees (tarriffs?).
Instead of directly generating Kicad, I chose to let it write "generators" in Python that output Kicad (S-Exps!) pcb and project - with mixed results - but I can generate a version with/without per-key Neopixels and customize certain parts of the board, including debug-stuff. I also _think_ this makes it easier to resume working on this in new sessions but not sure. Otoh, I never actually looked at the Python (vibed...) so it's still a lot of hard-coding.
I also landed on and used KicadRoutingTools after trying FreeRouting, combined with some minimal "auto-placing" code for larger ICs and such that are placed with certain constraints. My routing runs are way slower though - on the order of minutes.
I always wanted to try atopile - that might be a good match for AI/PCB work. The only thing that really held me back was the lack of routing.
> I always wanted to try atopile - that might be a good match for AI/PCB work.
Unfortunately all public open source work halted 6 months ago and they pivoted to a browser based IDE product to monetize: https://atopile.io/blog/atopile-v16
I followed them for a while before the AI craze started. It looked promising back then, but I wasnt willing to invest time and attention back then on something so new.
Atopile is ditching their cli. They're going the wrong direction.
It seems to me the first problem with AI in PCB design is that there isn't much in the way of machine friendly interfaces for the work. I watched Codex last night generate 3d renderings in Kicad to check the layout.
Anything that puts pressures on Nvidia's margins is going to have an outsized effect imo - margins declining from 75% to ~50% would undo all of its projected 2027 growth of 70%.
And what about the purported "AI frontier slowdown" - if frontier labs stop pushing better and better models, the main way to grow the pie is more users and that will eventually stop, too. Then providers would probably stop buying new hardware hand-over-fist and move to a slower depreciation-based replacement.
> 32B: Ranking among the top models in its class, 32B is our most powerful dense model, balancing capability, adaptability, and local deployability.
> 7B: The industry’s best-performing model under 10B combines strong software engineering and expert knowledge in a package small enough to run on a phone.
Sounds glib, but honestly ... more Ads :/