[add-sub-mul-div]

I cannot believe you people. The only difference between the status quo and the future is that the promotional messaging will be seamlessly and undisclosedly delivered with the conversational output. Ads are not going anywhere.

We can be forgiven for not seeing how social media was going to become weaponized against us, how streaming's promise of no ads was only temporary. There's no excuse for not seeing it coming this time.

[_factor]

Local AI searching for me and filtering the ads is my future. You can speak for yourself.

[const_cast]

Yes, I'm sure your "local AI" running on 8 gb of ram will be able to filter the ads added by the super computer AI just fine.

[maltelandwehr]

I don't think that is too unrealistic.

• Small models have performed really well in recent years. • Newer phones will have more RAM. • Private LLMs do not necessarily need to run locally.

[const_cast]

If it's not running locally, then by nature of what an LLM is it cannot be private.

LLMs cannot act on encrypted information - the server must have access to all your plaintext prompts. And you simply are forced to trust they don't keep them forever, but they probably do.

[bostik]

Just wait until the foundational models are all fed with increasing amounts of ads. After all, it will be the one remaining source of ongoing, if not human generated, at least human curated content. Ad exchanges will offer firehoses for a hefty fee, and the advertisers themselves will pay the exchanges extra to push their ads more frequently in there to gain a higher share of repetition.

The AI companies in turn will hoover in the deluge because they need something new to train their models with, embedding the ad copy deeply into the model itself.

Your local AI of the future will be just as ad-riddled.

[badsectoracula]

> Just wait until the foundational models are all fed with increasing amounts of ads.

Which foundational models?

Not all model providers are in the ad business and while the chances of building a supercomputer in your basement to train such a model are zero, some of the companies that build such models aren't exactly huge. Mistral for example is (according to Wikipedia) 150 people - this means that a company that can make their own model from scratch doesn't need to be some giant corporation. Which in turn means that it is possible for new companies to pop up, if there is a need for them - in this case, if some models become ad-infested, chances are other models will use their ad-free status as a feature.

And this is assuming only companies make such models. But some days ago i was reading here in HN about a new foundational model being trained by ETH Zurich and somehow i doubt a public university will inject ads in it.

[LegionMammal978]

Yep, 'dataset poisoning' is what I like to call it. If all newly-trained models mostly reflect the Internet, then the strategy is to shove your ad copy into such a large fraction of the Internet that models will accept it as true. Like today's SEO slop, but turned up to 11, since it's not even aiming for human clicks.