Comment by hattmall

17 days ago

The really strange thing is that so much of it doesn't work. Like I get that the SOTA models perform some tasks quite well and have some real value. But the AI being implemented in every corner creates a lot of really bad results. The Shopify code assistant will completely wreck your site and basically gets nothing correct. It will write 100 lines to change a color of a single DIV. The Amazon product Q&A will give you wrong information more frequently than not.

In what mind frame is it logical or necessary to put these extremely poorly functioning products in to the wild?

9 comments

hattmall

Reply
TeMPOraL  17 days ago

It's a desperate attempt at staying relevant, even if most of those companies don't realize it yet. Because of its general-purpose nature, AI subsumes products. Most software products that try to "implement AI in every corner" would, from the user's POV, be more useful if they became tools for ChatGPT/Claude/Gemini.

People's goals are rarely limited to just one software product, and products are basically defined as a bag of tools glued with UI, that work together but don't interoperate much with anything else. That boundary drawn around a bunch of software utilities, is given a name and a fancy logo, and sold or used to charge people rent. That's software products. But LLMs want to flip that around - they're good at gluing things, so embedding one within a product is just a waste of model capabilities, and actually makes the product boundary more apparent and annoying.

Or in short: consider Copilot in Microsoft Word, vs. "Generate Word Document" plugin/tool for a general LLM interface (whether Gemini webapp or Claude Code or something like TypingMind). The former is just an LLM locked in a box, barely able to output some text without refusing or claiming it can't do it. The latter is a general-purpose tool that can search the web for you, scrap some sites and run data analysis on results (writing its own code for this), talk results over with you, cross-reference with other sources, and then generate you a pretty Word document with formatting and images.

This is, btw., a real example. I used a Word document generator with TypingMind and GPT-4 via API, and it was more usable over a year ago than Copilot is even now. Partly because Copilot is just broken, but mostly because the LLM can do lots of things other than writing text in Word.

Point being, AI is eroding the notion of software product as something you sell/rent, which threatens just about the entire software industry :).

  • mark_l_watson  16 days ago

    I have been enjoying reading this thread, but with some irony: sure the email spams pushing their Lumo LLM private chatbot were a mistake, and I bet they stop doing that fast.

    The irony is that Lumo is a separate product, not really tied to the rest of their products except for a common login. Lumo works fine for the simple quality of life search and question answering stuff.

    Off topic, but have you tried avoiding the big corporate LLM providers and run local models? The small models just keep getting better and I find it fun and satisfying to do as much as I can locally.

Sharlin  17 days ago

"It's difficult to get a man to understand something when his salary depends upon his not understanding it."

In this case, the thing that's difficult to understand is "AI in everything is shit and nobody wants it."

pluralmonad  16 days ago

Saw an AI generated product feature list on walmart's site that listed a stainless steel rack as microwaveable. If someone can sue mcdonalds for hot coffee, I imagine someone burning their house down while microwaving steel probably could sue too. Intelligence of the plaintiff not withstanding.

  • Terr_  16 days ago

    > while microwaving steel

    There actually are microwave-safe steel objects, it depends on their shapes and conductive paths.

    After all, the whole inner-box is already a metal surface being blasted by the microwaves that come in through a small hole...