raynr.bsky.social

This is the wrong take. The latency is going to be a measure of network latency and not time to physically type. It’s unlikely that the North Koreans who can get jobs in Big Tech are awful typists. But they ARE a few thousand miles away, and it takes time for the keystroke to travel that far.

This is such a bad take I don’t even know where to start, it I’ll try. 1. I’m glad Google isn’t vibe coding the worlds most popular browser 2. There will be room for professional devs doing pro dev stuff for awhile But this is like Access or VBA.

Will this cause problems? Of course. I’m still supporting Access apps that should have been turned into real apps a decade ago. But those Access apps also let companies create things and scale when a real dev was not available

This is such a bad take I don’t even know where to start, it I’ll try. 1. I’m glad Google isn’t vibe coding the worlds most popular browser 2. There will be room for professional devs doing pro dev stuff for awhile But this is like Access or VBA.

Using @simonwillison.net 's llm tool in real life. I am NOT a mongodb expert, so when a service we use internally failed due to DB errors I wasn't sure where to start. Solution: cat dclogs.txt |grep "chat-mongodb"|tail -n 100|llm -c -m 'gemini-2.5-pro' 'what does this tell you?`

Would I have gotten there on my own? Eventually, sure. But llm + gemini turned this into a 15 minute fix instead of ~2ish hours.

Using @simonwillison.net 's llm tool in real life. I am NOT a mongodb expert, so when a service we use internally failed due to DB errors I wasn't sure where to start. Solution: cat dclogs.txt |grep "chat-mongodb"|tail -n 100|llm -c -m 'gemini-2.5-pro' 'what does this tell you?`

"The most significant [limitations] to [MCP servers]... is in token usage... Almost everything I might achieve with an MCP can be handled by a CLI tool instead." - @simonwillison.net I think you're doing it wrong 🙃 1/

The Firecrawl MCP server is another example that shines as a token-saver in a subtask. "You look stuck on XYZ, please stop and use a subagent to search the web for more info on library ABC" COULD python do that? With a Bing/Google library, sure, & get back 10 pages of results looking for 1 page. 7/

"The most significant [limitations] to [MCP servers]... is in token usage... Almost everything I might achieve with an MCP can be handled by a CLI tool instead." - @simonwillison.net I think you're doing it wrong 🙃 1/

Not an MCP but demonstrative of #2, a lot of my claude[.]md files have this: You can use the bash tool 'llm' to ask questions of the codebase or the docs, this is AMAZING for working with the massive API doc. Ex: ``` cat MyAPIDocs.html|llm "What are the properties for the contacts entity?" ``` 5/

"The most significant [limitations] to [MCP servers]... is in token usage... Almost everything I might achieve with an MCP can be handled by a CLI tool instead." - @simonwillison.net I think you're doing it wrong 🙃 1/

2. Shift your thinking as the main LLM convo as the orchestrator, not the primary agent. At least, to be more-orchestrator like. 3/

"The most significant [limitations] to [MCP servers]... is in token usage... Almost everything I might achieve with an MCP can be handled by a CLI tool instead." - @simonwillison.net I think you're doing it wrong 🙃 1/

@davidfrenchjag.bsky.social I’d love to see an op-ed that expands your theory on the interaction of Rom 13 and Democracy vs Emperor, as outlined in “Why MAGA Evangelicals Can Cheer Love and Hate at the Same Time” Something that colors in the details, esp supporting that the idea that…

#ChatGPT lies about its user-agent in agent mode. In regular mode, it will return Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko); compatible; ChatGPT‑User/1.0; +https://openai.com/bot In "agent" mode it returns YOUR user-agent. #openai #ai

#ChatGPT lies about its user-agent in agent mode. In regular mode, it will return Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko); compatible; ChatGPT‑User/1.0; +https://openai.com/bot In "agent" mode it returns YOUR user-agent. #openai #ai

The worth of social media is determined in large part by the network effect: the value or marginal users is determined exponential, not linear. Value = users^y For #ai it will be access to information. Value = information_sources_available^y

Related: it baffles me why ChatGPT will search the web so willingly but 4o et al require so much more prompting to make use of the tools well.

The worth of social media is determined in large part by the network effect: the value or marginal users is determined exponential, not linear. Value = users^y For #ai it will be access to information. Value = information_sources_available^y

Not all sources should be available at any given moment, but it should be easy for the LLM or the user to quickly find the best source for a given task and add it to the context.

The worth of social media is determined in large part by the network effect: the value or marginal users is determined exponential, not linear. Value = users^y For #ai it will be access to information. Value = information_sources_available^y

We need to stop calling LLMs non-deterministic, it’s not true. AI systems built on LLMs appear stochastic (random) because we build them that way. If you run the exact same parameters thru the same model multiple times you will get the same result*. 1/x

My humble suggestion: we refer to AI systems as random, but AI models (including LLMs) deterministic. 6/x

We need to stop calling LLMs non-deterministic, it’s not true. AI systems built on LLMs appear stochastic (random) because we build them that way. If you run the exact same parameters thru the same model multiple times you will get the same result*. 1/x

So why do we call them random? Because when you interact with ChatGPT — an app that *includes* the LLM but is not itself the LLM — randomness is injected into the chat. It wouldn’t be a good user experience if the chatbot responded the exact same way every time. 4/x

We need to stop calling LLMs non-deterministic, it’s not true. AI systems built on LLMs appear stochastic (random) because we build them that way. If you run the exact same parameters thru the same model multiple times you will get the same result*. 1/x

This matters, (1) to be accurate and (2) to disallow AI companies from disclaiming all responsibility. “We tried to stop it praising Hitler but it’s random, we can’t do anymore.” 2/x