davidghowell

Joined 1 year ago

Karma

Submissions

Member since

Jul 9

Recent submissions

"rejects honest contributors while rewarding anyone willing to conceal their tools" is exactly the gaming that blanket bans lead to.

A common worry cited about publishing a lot of AI generated writing is that proliferation of it on the open internet makes it available to be used in future AI model training data. The concern is that recursive training will cause some kind of regression-to-the-mean.

I'm not sure it is strictly true. ML training approaches often use augmentation to enrich the data and get better results. In computer vision, for example, rotating or skewing the same image gives more training examples and leads to better outcomes.

I also agree with another main point you make. You are the one who decided to write the article, you argued the points and guided the tone. AI was just the assistant. The blog post might be AI generated, but AI of itself has no agency. The very word "agentic" is a misnomer in that sense.

<Facetious/> Now, given you subscribe to the Sapolsky idea in "Determined" that there is no free will, it then follows that you didn't have any true agency here either. So who really directed the blog post to be written?

I'm poking fun, but in all seriousness I don't think the answer to that question really matters. You made good points, you used the AI to clear it up and add structure to the argument and correct all the grammar, and the outcome is the product of the human+AI cyborg chimera that we are inevitably becoming.

Comment · 5 months ago

That's picking up all the "bunsafety" references in there :P

Comment · 1 year ago

The point seems to be to just use 1 instead of all 3 runtimes. OP needs the Deno sandbox and security features so that is the deciding factor.

Comment · 1 year ago

The security issues are probably orthogonal in the way most people install and use these MCPs, but the article mentions Cloudflare's "code-mode" running in v8 isolate sandboxes and calling rpc bindings that are pre-authed, no API keys or open slather internet access required, see: https://blog.cloudflare.com/code-mode/#running-code-in-a-san... This is at least interesting, possibly even novel.

"Wasting context window to understand and re-implement each integration is why MCPs exist" does seem to be exactly the point. Pointing the LLM at a swagger/openAPI spec and expecting it to write a good integration might work, but gets old after the first time. Swagger docs focus on what mostly, LLMs work better knowing why as well.

And,why not just use a locally installed cli rather than an MCP? You need to have one for a start, and use-cases (chained calls) are more valuable that atomic tool calls.

There is more behind the motivation for MCP and "tool calling" ability generally with LLMs. This motivation seems less and less relevant, but back when reasoning and chain-of-thought were newly being implemented, and people were more wary of yolo modes, the ability for an LLM harness to decide to call a tool and gather JIT context was a game changer. A dynamic RAG alternative. MCP might not be the optimal solution long term. For example, the way that claude has been trained to use the gh cli, and to work with git generally is much more helpful than either having to set up a git MCP or getting it to feel its way around git --help or man pages to, as you said "re-implement each integration" from scratch every time.