alextillman

Right, but you still don't know that the big model has exactly the same "jargon" handling as you do. Plus it means that you can't use smaller models not trained on that jargon. That seems like a really limiting black box to place at the heart of a system.

This doesn't resonate:

Consider this prompt:

Make sure that, whenever a job runs, we can tell where it started from, who triggered it, what settings were passed in, and what changed later. In the future, if something breaks, we want to be able to trace it back and understand what happened.

That is fine.

But an expert can say:

Jobs should persist provenance metadata."

That only works if the model is trained, specifically the way you want, to understand that second sentence. If not, any model could work with the first sentence, but not with the second.

You've crated a need for expert training at the model level (which is insanely expensive to create and maintain) rather than accessible natural language discussions that work on any model and understood by anyone. Denser isn't "better" because the words have more power.

What's old is new again. The solution RSS offered was structure for an otherwise unstructured challenge (trying to figure out updates on a site). That value grew exponentially when connected to AI (providing the signals of when do I need to look at this site/podcast again). Smart marketing.

Nothing mentioned here is specific to AI. EVERY SaaS service is designed to monetize engagement rather than provide the nest service. These services typically use dark patterns learned from casinos to know how to try and trap users and make them spend more.

"So the next time you get AI to generate something for you, be wary of the early output. Be critical, demanding, and opinionated on what needs to be improved. An amateur considers the 3rd iteration good enough; a pro takes it to the 30th or 300th and probably starts over a few times."

You can test this with any platform, simply type "improve this response" after any result and see what you get back (no thinking credits, nothing fancy just three words and a few seconds)

This does not seem remotely true in my experience. This seems like more “Bad prompt generated entirely predictable garbage results” rather than “We rigorously tested different scenarios here, and AI seems to ‘ignore' new concepts”

Interestingly they did not share what prompts they used to run any of this – highly suspect.

And those US hosted versions of DeepSeek etc. are still far cheaper than the big (frontier) US companies... just not as cheap as the Chinese hosted sites.

Makes sense in terms of public information. But the thing is if you use a foreign-hosted tech, you don’t know what is doing with ALL data you signed up with. You can use Chinese models hosted on US servers (like a together.ai – there are others). You’ll pay more than the native Chinese app sites, but will know that you aren’t feeding personal data into an ominous black hole. One simple example is to think what “dossier” they could build with your credit card data and search data. Nothing good can come from that.

This misses out on what AI (when used properly) can do. If I build my own CRM that uses off the shelf AI (collects form data, searches for relevant social media and company information, builds a dossier on the person etc – all fairly simple things to do with the right tools) and it does ALL the things that Salesforce does (and more) at a fraction of the price, I no longer need Salesforce. And if improving my AI-based program requires no code, just more refined instruction sets, I don’t need devs. If I do the same thing for email, marketing, social posting etc. I can then run a business with tools that I tell how they should work, and that do a better job than those in the SaaS marketplace. At that point, I am not just replacing jobs, I am replacing companies.

Maybe your job isn’t at risk of being replaced, but your entire organization is.

This is like a major highway. They get overloaded at rush hour, so they build lanes to accommodate peak traffic. But travel any highway at 2AM (outside of LA/NYC sized metropolises which can have rush hour at any time) and you get an entire highway to yourself. The problem is rush hour. If more AI requests could be load balanced to run overnight, we’d solve a large part of the current pressure. But everyone demands search engine-like response times, even if overnight wouldn’t be a problem.

This is because the current AI approach relies on AI to be a glorified search engine – know everything about everything requiring enormous, ever growing models, and demanding search-engine like near instant responses requiring bigger more complex chips and sprawling data centers to run them in. This leads to a loop demanding ever bigger models, updated at a more and more expensive cost, and chipsets that become much more expensive to deploy.

If you move those things to software and utilize tools that are cheap at scale (databases, web search etc.) the hardware arms race ends and the price becomes sustainable. With the right tools preparing dynamic context for a conversation, models are used for their reasoning and not for their knowledge. And waiting even a minute or two for a model to prepare a response, evaluate it, and iterate to improve quality makes a huge difference.

The AI economic bubble is not because AI is useless. There was a paper recently that highlighted how AI is diagnosing pancreatic cancer patients years before humans do. This is not a one off (and not limited to science). Those who are getting these types of results across many domains understand the potential value (like the people who get a two year head start on their treatments).

The problem is that the entire market right now is set up for resolving things with hardware instead of software. Bigger models, faster chips, data center sprawl etc. Clearly that is not sustainable. But if you use AI inference for reasoning only, and traditional software like databases, web searches, etc. for the things they do well (at a fraction of the cost) the economics flip.

Trying to force expensive training updates and ever larger models that know who the first baseman was for the winning 1939 world series time is useless and where the cost-to-value ratio is broken. Pull what you need from a search api, and send the results to an LLM to analyze. No fine tuning, no super big chips and even small models can produce good results.