gharman

I’ve been doing “marketing” experiments with adjusting MCP descriptions to try to optimize selection of a particular service. Not just maximizing “clicks” but optimizing type 1 and type 2 errors - choose this service when it would fit the task, and move along when it won’t.

What have I found? The frontier models (my main targets) usually don’t search for MCP tools at all from catalogs or even the providers’ own “marketplaces” if you haven’t specifically installed or directed the model to do so. The models have a strong bias towards attempting to DIY a solution instead of searching (for tools).

They will happily search the web for domain-related information. So the answer to the headline question seems to be: nope because the models don’t do that. But it could be if the next generation were tuned differently to have that preference.

Yes! Paraphrasing my childhood:

“Mom, I’m bored will you buy this a new computer game?” “Lol no! Go outside.” “Well crap. I guess I’ll have to write my own game.”

> My new suspicion is now that they didn’t got drastically smarter, but they got trained on the user input on the previous generations.

Yes - that’s one of the ways in which they got smarter. More data more better.

I presume also lessons learned re training procedures, algorithmic advances, but data is a primary lever.

Any (non?)digital media? I would be surprised not to see this exact dynamic on BJJ Fanatics. So easy to pick up instructionals with the constant stream of sales and discounts. Not so easy to spend the hours it takes to watch and digest one of those.

Commercial libraries = private library where you pay a membership fee? I’ve not really heard of that here outside of arguably university libraries and similar.

But SF public library has a couple dozen branches or more?

> (as an aside: bland food won't work well on someone who lost their sense of smell from long covid, and prison populations were heavily affected)

The consistent texture of the Nutraloaf may “help” here. My spouse lost her sense of smell (for other reasons than Covid) years ago and now relies heavily on textures for food enjoyment.

IDK when I hire, the first list (as might be directly relevant to the job) is checking a box, but the second list is what catches my attention.

Agreed there is a semantic argument here. I don’t think anyone is anthropomorphizing LLMs. I’ve had to adopt “objective” with one customer instead of “task” (my preferred default term) because their industry overloads “task” with something else. Now that said, as you point out, the real objective is next token prediction not “help the user with a truth bomb.” and it just happens to have emergent accidental usefulness (I think a natural bias from the training data).

They (chatgpt et al) do move closer to “clear and coherent” by adding extra layers such as a beam search on LLM outputs. Good to remember that ChatGPT et al are products, not bare metal LLMs.

My elderly father still uses this as his primary gateway to the internet and email. For all intents and purposes, this is the internet to him.

Looks like Scholar Turbo has GPT-4 and Chatpdf is still on the waitlist. But yeah, these chat-with-your-doc webapps are popping up out of the woodwork.

My long-horizon strategy has been the kinds of roles I take on. A product-focused role lets you go deep into a narrow problem area, whereas a customer-facing consulting/services role forces you to go wide, learning a new area every time. Over the years I’ve used time in the latter to select what to go deep into.

Indeed! I built a system just last year with - count em - three parsers to deal with PDF table extraction, including one built on TableTransformer. And then when GPT4 came out I just copy pasted a PDF into it as-is and darned if it didn’t do at least as good a job.

Now I can’t do this in earnest because of document privacy issues but I’ve diving down the rabbit hole of how small can we go and still get decent results. Spoiler: gpt2 is too small. :-)

> it’s 30+ minute charge time. As great as electric is, an extra hour in the car on a road-trip with 3 kids is a tough sell.

I dissent! Just did a 2-day (each way) road trip with young children and those charging breaks are pretty great and well-spaced for letting them run around and blow off steam. It’s not such a bad idea for the adults either…

IDK I’d say high school. I took a semester welding class at community college and we only spent 3 sessions on MIG (what this appears to be) before moving on to other processes. Most of the learning was around dealing with PPC, listening for the sizzle, identifying a good bead and common mistakes. This seems expensive for a one-class tool, especially since they still need the regular welders to complete the training.

You can choose to program functionally in Java or Python! I do this enough that seeing object oriented Python irritates me. I see this as more of a stylistic choice than forcing new languages.