
Robert MacCloy
@rbm
Followers
397
Following
2K
Media
24
Statuses
286
Programmer-type. Distributed systems, developer tools, internet plumbing. Early at @Quantcast, @HearsaySystems, others. Building a new thing @ScrunchAI
Los Angeles, CA
Joined November 2006
The two cases where you can reasonably try to understand LLM behavior by _asking_ an LLM are 1) interrogating the system prompt and 2) interrogating reasoning traces (or other pre-registered plans). Otherwise, at best, you're getting a plausible explanation, not the actual cause.
@zeynep That's the output of the model (I think Claude Sonnet 4). It should not be taken literally. All explanations of its conduct (and internal states) are unreliable. Even more so when it is a follow up prompt. It's just looking back at the convo and weaving a coherent narrative.
1
0
2
RT @lefthanddraft: @zeynep That's the output of the model (I think Claude Sonnet 4). It should not be taken literally. All explanations of….
0
1
0
RT @vsreekanti: We're generally focused on where AI fits into each of our businesses, but we're not paying as much attention to what happen….
0
2
0
people either: .- don't want to browse your website at all.- _really do_ want to browse your website. IMO, few cases where someone wants to browse your website, but. just a bit less. For 30 years browsing has been synonymous with "doing stuff on the internet". times change.
I think it's worth contextualizing the concept of "AI Browsers" against the behaviors people might want to automate. Looking at top websites by visits and time spent, we can quickly see these browsers are mostly search and shopping plays. Not a small slice!
0
0
0
RT @DylanByers: “Google is shifting from being a search engine to an answer engine… We have to develop new strategies.”.
0
2
0
RT @tanayj: ChatGPT is now a global top 5 website by traffic. Also the only one in the top 10 still on the rise. (via @Similarweb) https://….
0
50
0