73% of AI Startups Are Just Prompt Engineering

It's up to the domain experts and me to understand where giving it data will tone down the hallucinative nonsense an LLM puts out, and where we should not give data because we need the problem solving skills of the LLM itself. A similar process is for tool-use, which in our case are pre-selected Python scripts that it is allowed to run.

- write an evaluation pipeline to automate quality testing

- add a query rewriting step to explore more options during search

- add hybrid BM-25+vector search with proper rank fusion

- tune all the hyperparameters for best results (like weight bias for bm25 vs. vector, how many documents to retrieve for analysis, how to chunk documents based on semantics)

- parallelize the search pipeline to decrease wait times

- add moderation

- add a reranker to find best candidates

- add background embedding calculation of user documents

- lots of failure cases to iron out so that the prompt worked for most cases

There's no "just give LLM all the data", it's more complex than that, especially if you want best results and also full control of data (we run all of that using open source models because user data is under NDA)

Not sure if we called it engineering ten years ago.

It seems right now like there is a tradeoff between creativity and factuality, with creative models being good at writing and chatting, and factuality models being good at engineering and math.

It why we are getting these specific -code models.

But cloud services run in... the cloud. It's as big as you need it to be. My cloud service can have as many backing services as I want. I can switch them whenever I want. Consumers don't care.

"One model that can do everything for you" is a nice story for the hyper scalers because only companies of their size can pull that off. But I don't think the smartphone analogy holds. The convenience in that world is for the the developers of user-facing apps. Maybe some will want to use an everything model. But plenty will try something specialized. I expect the winner to be determined by which performs better. Developers aren't constrained by size or number of pockets.

I don’t think that’s the expectation set by “everyone else” in the AI space, even if it arguably is for OpenAI (which has always, at least publicly, had something of a focus on eventual omnicapable superintelligence.) I think Google Antigravity is evidence of this: there’s a main, user selected coding model, but regardless of which coding model is used, there are specialized models used for browser interaction and image generation. While more and more capabilities are at least tolerably supported by the big general purpose models, the range of specialized models seems to be increasing rather than decreasing, and seems likely that, for conplex efforts, combining a general purpose model with a set of focussed, task-specific models will be a useful approach for the forseeable future.

Foundational models are not great for many specific tasks. Assuming that one architecture will eventually work for everything is like saying that x86/amd64/ARM will be all we ever need for processors.

Which would be a major security hole. And sure, lots of startups have major security holes, but not enough that he could come up with these BS statistics.

I'm a little dismayed at how high up this has been voted given the data is guaranteed to be made up.

They claim to have found that.

You mean, except for explaining what he's doing 4-5 times? He was literally repeating himself restating it. Half the article is about the various indicators he used. THERE'S EXAMPLES OF THEM.

There's this bit:

> Monitored their network traffic for 60-second sessions

> Decompiled and analyzed their JavaScript bundles

Also there's this whole explanation:

> The giveaways when I monitored outbound traffic:

> Requests to api.openai.com every time a user interacted with their "AI"

> Request headers containing OpenAI-Organization identifiers

> Response times matching OpenAI’s API latency patterns (150–400ms for most queries)

> Token usage patterns identical to GPT-4’s pricing tiers

> Characteristic exponential backoff on rate limits (OpenAI’s signature pattern)

Also there's these bits:

> The Methodology (Free on GitHub next week):

> - The complete scraping infrastructure

> - API fingerprinting techniques

> - Response time patterns for every major AI AP

One time he even repeats himself by stating what he's doing as playwright pseudocode, in case plain English isn't enough.

This was also really funny:

> One company’s “revolutionary natural language understanding engine” was literally this: [clientside code with prompt + direct openai API call].

And there's also this bit at the end of the article:

> The truth is just an F12 away.

There's more because LITERALLY HALF THE ARTICLE IS HIM DOING THE THING YOU COMPLAIN HE DIDN'T DO.

In case it's still not clear, he was capturing local traffic while automating with playwright as well as analyzing clientside JS.

This makes literally no sense. Why would any companies (let alone most of them) contract with this guy who seems hell bent on exposing them all.

The article is simple made up, most likely by an LLM.

> GPT‑5 is a unified system with a smart, efficient model that answers most questions...

It doesn't say it's a new model, but is a new routing design.

At what point can you claim that you did "it"?

Do you have to use an open source model instead of an API? Do you have to fine tune it? How much do you need to? Do you have to create synthetic data for training? Do you have to gather your own data? Do you need to train from scratch? Do you need to come up with a novel architecture?

10 years ago if you gathered some data and trained a linear model to determine the likelihood your client would default on their loan and used that to decide how much, if any, to loan them- you're absolutely doing "actual AI"

---

Any other software you could ask all the same questions but with using a high level language, frameworks, dependencies, hiring consultants / firm, using an LLM, no-code, etc.

At what point does outsourcing some portion of the end product become no longer doing the thing?

Btw, the so-called AI devs or model developers are "users" of the databases and all the underlying layers of the stack.

The thing that annoys me is when clearly non-AI companies try to brand themselves as AI: like how Long Island Iced Tea tried to brand themselves as a blockchain company or WeWork tried to brand themselves as a tech company.

If we’re complaining about AI startups not building their own in house LLMs, that really just seems like people who are not in the arena criticizing those who are.

All of these are things that will need to be solved long-term in the model itself though, at least if the AI bubble needs to be kept alive. And solving those things would in fact materially improve all sorts of benchmarks, so there's an incentive for frontier labs to do it.

I think this is why you have the back-forth pattern that GP mentioned. You start with a single model doing everything. Then you find all sorts of gaps that you start to plug ad-hoc, and decide that breaking it into subagents might help fix things. This works for a while but then you realize that you lose out on the flexibility of a single-model having access to the entire context, so you starting trying to improve communication between subagents. But then a new model drops that fixes a lot of the things you originally had to workaround, so you go back to a single-model setup. Rinse and repeat. It's a great VC- bubble funded employment program though.

I think the reason of the recent pivot is to “keep the human in the loop” more. The current thinking is they tried to remove the human too much and were getting bad results. So now they just want to make the interaction faster and let the human be more involved like how we (developers) use Claude code or copilot by checking every interaction and nudging it towards the right/desired answer.

I got the sense that management isn’t taking it well though. Just this Friday they gave a demo of the new POC where the LLM is just suggesting things and frequently asking for permissions and where to go next and expecting the user to interact with it a lot more than the one-shot approach before (which I do think is likely to yield better results tbh) but the main reaction was “this seems like a massive step backward”