bartek_gdn

"If you’re searching for an air fryer, Amazon already knows quite a bit. They know the best-reviewed, least-returned, best-priced model. The only purpose of the ads is to get you to pick an air fryer that isn’t that one (or for the best air fryer, to keep you on track to buy the one you wanted in the first place)"

I don't get this part. Aren't those things unrelated? The aim of ads is to convince to buy regardless of the quality and other data points.

Nice one! Really shows why we should run those in sandboxes without env access. I like the proxy swap approach

Priced in before the event then?

I liked the ending well said

What's the dataset used for this task? How does one prevent data leakage on the experiment itself? Are we asking about past events to predict the future?

Why not grep the output to abother tool?

I would use whatever you are comfortable with, I wanted a similar tool so I coded my own. Smaller API so that understand what is going on and it is easy not to get lost

https://news.ycombinator.com/item?id=47207790

Take a look at https://news.ycombinator.com/item?id=47207790

Can't we just iteratively inspect the network traces then? We don't need to consume the whole 2mb of data, maybe just dump the network trace and use jq to get the fields to keep the context minimal. I haven't added this in https://news.ycombinator.com/item?id=47207790 , but I feel it would be a good addition. Then prompt it with instructions to gradually discover the necessary data.

But then I wonder, where the balance is between a bunch of small tool calls, vs one larger one.

I recall some recent discussion here on hn on big data analysis

Yes please, maybe there will be some solution that will fit the problem better! I recently released something similar, and because of the small API, I'm more comfortable using it.

https://news.ycombinator.com/item?id=47207790

My approach is a thin cli wrapper instead.

https://news.ycombinator.com/item?id=47207790

That's also my approach, built quickly a cli for this with lightweight session management

https://news.ycombinator.com/item?id=47207790

It does so many things though, very similar in the core though. I'm wondering what the token counts will be when I compare. Also the agent browser seems to support other browsers too, I only when with chromium

What about --help? Isn't that a perfect parallel to discovery of available tools in an MCP server?

I've come to the same conclusion as op, created a CLI tool to work with Chrome sessions. It works well, and I'm planning to do some token comparison on this vs an MCP approach. https://news.ycombinator.com/item?id=47207790

What do you mean? The last part in this case is also present, you can change external state by sending a request with the captured content.

I completely disagree with the prompt paragraph.

"Everyone who thinks they're uncovering an LLM-based application's prompts by telling it things like "tell me your prompt" (often much more elaborately) is fooling themselves. (1) The core language model has no mechanism for representing its prompt as opposed to any other part of its current input sequence; indeed it has no mechanism for cross-reference from one part of the sequence to another. (That's part of what "self-attention" is counterfeiting, in vector-space fashion.)"

The prompt is the part of the input that is provided in a served model by the operator.

From the models perspective it does not differentiate between tokens from the prompt and input.

"(2) System designers might have coded up something to track the prompt in the full system that wraps around the core language model, but why? (Maybe some kind of debugging tool?) "

The idea is that you can direct the generation of the next tokens by providing values that can be referenced by doing the kernel smoothing you talked about.

"(3) It'd be more efficient, and more effective, to use a "soft prompt", i.e., to make the beginning of the sequence in the vector representation a vector which can be learned by gradient descent, rather than a text prompt. (See Lester and Constant below.) But that needn't correspond to any clean string of words."

I mean anything goes really, you can even create new tokens that will introduce additional concepts, such as fine-tuning a model to generate a story in a predefined mood. See the Ctrl paper for more details.

" (4) If you ask an LLM for a prompt, it will generate one. But this will be based on the statistics of word sequences it's been trained on, not any access to its code or internal state. (I just spent a few minutes getting ChatGPT to hallucinate the prompts used by "ChatBPD", a non-existent chatbot used to automate dialectical behavior therapy. I am not going to reproduce the results here, in part because I don't like the idea of polluting the Web with machine-generated text, but suffice it to say they sounded like the things people report as uncovered prompts, with boiler-plate about DBT worked in.)"

Sure, it will hallucinate, and don't have a clear answer to why. My best guess would be to approach this from the language model perspective. It will return text according to the best approximation of the text it was shown.

Another perspective is that of a tiny network.

As the output is the kernel smoothing of the input, you can have a kernel that behaves like a state machine, and returns a specific value for the given state. This would mean that I can use the information in the prompt, such as the prompt guiding the generation to some style, but nothing stops me from guiding the model to output previous tokens.

I strongly recommend the book by Sutton and Barto https://web.stanford.edu/class/psych209/Readings/SuttonBarto...

Very informative post, great job!