Microsoft's Plan To Fix the Web: Letting Every Website Run AI Search for Cheap (theverge.com) 22
Microsoft has announced NLWeb, an open protocol designed to democratize AI-powered search capabilities for websites and apps. Developed by Microsoft technical fellow Ramanathan V. Guha, who previously created RSS and Schema.org, NLWeb allows site owners to implement ChatGPT-style natural language search with minimal code. The protocol enables websites to process complex queries like "spicy and crunchy appetizers for Diwali" or "jackets warm enough for Quebec," requiring only an AI model, some code, and the site's own data.
During his demonstration to news outlet The Verge, Guha showed how NLWeb remembers user preferences, such as dietary restrictions, for future interactions. "It's a protocol, and the protocol is a way of asking a natural-language question, and the answer comes back in structured form," explained Guha, who argues the approach is significantly cheaper than traditional search methods that require extensive web crawling and indexing. Microsoft is partnering with publishers and companies including TripAdvisor, Eventbrite, and Shopify to implement NLWeb, though Guha acknowledges the challenge of achieving widespread adoption in a web that historically tends toward centralization.
During his demonstration to news outlet The Verge, Guha showed how NLWeb remembers user preferences, such as dietary restrictions, for future interactions. "It's a protocol, and the protocol is a way of asking a natural-language question, and the answer comes back in structured form," explained Guha, who argues the approach is significantly cheaper than traditional search methods that require extensive web crawling and indexing. Microsoft is partnering with publishers and companies including TripAdvisor, Eventbrite, and Shopify to implement NLWeb, though Guha acknowledges the challenge of achieving widespread adoption in a web that historically tends toward centralization.
Attack vector (Score:3, Insightful)
Another security nightmare - securing your llm against malicious prompts attempting to DDoS it by getting it to consume resources, or to divulge privileged information it has access to. This is like putting a sociopathic teenager in charge of customer support (aka the DOGE methodology)
Re: (Score:2)
Not following your reasoning here. Are you implying something about using the AI-enhanced websites as some kind of weapon in a form of jujitsu? If so, I'm having trouble seeing how the direction of causality is supposed to work...
But I am looking forward to Funny. Story sure seems ripe for jokes. I was just checking for some early example of "What could possibly go wrong?" and this FP seems to be falling short.
Slowing the decay and end of existing web sites (Score:2)
Conjecture: Lots of web sites have disappeared (pandemic) and will disappear (boomers retiring and closing businesses, AI generated answers). Microsoft, google, and other big tech will need to slow the decline because AI slop will reduce original human generated content to a small fraction of the web and drive away most users.
Conjecture 2: Social media and social media adjacent (LinkedIn) sites will lose users as long-time users age out and getting new users who engage with the community becomes more dif
Re: (Score:2)
Thanks for the thoughts. Some worth comment, but discussion basically expired as Slashdot measures time, so I'll just think about them for now...
What does NL Web have to say about NLWeb ? (Score:4, Interesting)
AI all the way down (Score:1)
Re: (Score:2)
Currently there is too much development. There are like four new models every week (speaking about open weights, you bet there are a lot more that are not shared). Any candidate you could choose will be outdated by the end of the year.
It sounds like it's just a local search tool (Score:4, Insightful)
If you're actually running the searches and the model off of Microsoft systems then odds are good they've got your data and they're going to use it to train their own AI models which you probably don't want because you're giving up a ton of incredibly valuable data you could be selling to Microsoft for profit.
On the other hand if you're running it on your own servers AI models require a shitload of computer power and well, you're going to have to pay for that.
I guess in exchange for that you get an AI chatbot that you can use to sell crap to people.
The actual use case of an AI chatbot for it quickly accessing content though is terrible. You want users to have to browse around your website because 99% of the web business model is having people browse and then you get the throw ads at them on each page. If they can very very quickly find what they're looking for they're going to come in, get it, leave and you don't need much revenue because you don't have eyeballs.
That's a huge part of enshitification. You have websites designed to force you to engage for the maximum amount of time possible. On the other hand without that the entire content industry that generates a lot of the useful information on the internet kind of collapses.
Re: (Score:1)
If it is for find-in-page++ you don't need a large model. For summarizing and question answering to a text in the input you'll probably choose a tiny model fine-tuned for that tasks, and that can run on a raspberry pi. Have a look how good (for its size) Qwen3 0.6B is. And I think people run up to 8B models with wasm in the browser. So I'd say a good 1B model as Q4 quant would work on most devices people use today.
Re:It sounds like it's just a local search tool (Score:5, Informative)
It's all about ads and tracking. The search history, answers delivered, all of that will be the crux of still other training models, so that much can be inferred by every hook possible.
Have AI do shopping for you? You plainly know who's going to be at the head of the list-- someone who paid for it. Recommend travel, destinations, therapists, tires, games, the list is infinite, will be monetized, tracked for effectiveness and will learn how to sell exactly to YOU.
For a price, with micro-payments to $search_API.
Re: (Score:2)
The problem is that besides LLM use for replacing workers I don't see a lot of value in it.
If you make content for a living AI basically takes your content and regurgitates it, often with only the bare minimum pieces of it that people want. It means that as a user you quickly get to the information you want, assuming AI is working like it's supposed to, but it means that you're engagement statistics and with i
Re: (Score:2)
The Ad AI business is different than the GoogleAds mindset. It makes money in different ways, and still needs ad composition, placement, review, sales people, marketing, infrastructure, an HR dept (perhaps for bots), etc. It's not a zero-overhead business.
Specialty AI will become a norm, e.g. like subject matter experts of a different era. In turn, these will be networked together in their own matrices. Those matrices will micro-monetize information, transactions, and placements, as well as tasks (compute).
Anyway mark my words AI will (Score:1)
Will doom us all sooner than later. As far as a way around that there really is only one and it's to change how we teach to focus on critical thinking and improving children's capabilities to the maximum they can reach. But that's it odds with cost cutting and a desire to create effective corporate robots as fast as possible and it's cheaply as possible like natsi Musk is doing with his team of h-1bs. It's too expensive to move a warehouse. Robots is devouring jobs.
Tech workers and customer service workers
Re: (Score:2)
I'm not quite as paranoid as you are, but we share similar concerns. However, this is about Microsoft and the post.
It's an API, one step towards both panopticon, making Microsoft loot, and also an early embryonic set of ideas not unlike how computing replaced human tasks 30-40 years ago. There were many exclamations of how those fun menial tasks (that paid actual money) might be automated. Those that went with the automation survived. It's Darwinian in that characteristic.
Now, faced with bots, AI, and other
Can't tell which egregious sin this is. (Score:5, Interesting)
Is it a way to bypass all possible security and scrape all information on web sites into Microsoft AI systems? Or is this a hoped-for future addiction that will feed a payment plan for any website dumb enough to use it today, to be monetized later? Or is it a better way to gather info for targeted advertising by scraping every site that adds this "open protocol?"
I mean, there's so many possible nefarious reasons for this to exist, I'm having a hard time coming up with a single possible positive for the end users and website owners. It all seems to point to, "Give us more direct hooks to your information, please. We needs dat data." And possibly to, "We'll monetize egregiously once we get you addicted."
Re: (Score:1)
MS's new NefariousGPT generates creative vendor-lock-in and data vacuuming ideas that took human MBA dolts years to devise.
MS's new motto: Be Evil Faster
Google is suing for motto swiping.
Fixing the web? (Score:3)
Reduce JavaScript to a small subset, limit the number of third party sites that can be used to load resources, do not allow scripts or cookies from third parties at all. Maybe also reduce HTML to a safe subset. Kinda what AMP could have been if it wasn't from Google.
AI search on websites is not required to fix anything. I am very open to the better uses of AI, but embedding it in every website sounds like one of the worst. But since Recall it looks like Microsoft is trying to find what the worst use-case for AI might be.
Pass... (Score:1)
Guha created RDF, not RSS. (Score:2, Informative)
> Ramanathan V. Guha, who previously created RSS
consulted about, yes. created: no.
to be fair: RDF was a much larger technical achievement and RSS was only a minor application of it. (until later, when it wasn't even that).
But AI is what is currently breaking the web. (Score:2)
Sounds like Semantic Web (Score:2)