Kara

lukewarm take: opus 5.5 still sucked at technical writing at launch, opus 5 just set the bar so low that people ate it up anyways

another real opus 5.5 example: "A rule that references an identity on a resource that never sends one cannot widen access" consider instead: "If the resource is not configured to send an identity but has a rule referencing one, the system fails closed, restricting access"

lukewarm take: opus 5.5 still sucked at technical writing at launch, opus 5 just set the bar so low that people ate it up anyways

granted, it's not like a human would never write the former example (I have seen similar before) but it's just unclear writing and not what I would expect from a skilled communicator who wants to avoid being misunderstood

lukewarm take: opus 5.5 still sucked at technical writing at launch, opus 5 just set the bar so low that people ate it up anyways

another real opus 5.5 example: "A rule that references an identity on a resource that never sends one cannot widen access" consider instead: "If the resource is not configured to send an identity but has a rule referencing one, the system fails closed, restricting access"

lukewarm take: opus 5.5 still sucked at technical writing at launch, opus 5 just set the bar so low that people ate it up anyways

perhaps informatively, GPT models seem to be the end-state of this failure mode, where rather than weird connective language it usually prefers no connective language at all, resulting in a string of self-contained facts with minimal throughline connecting them

lukewarm take: opus 5.5 still sucked at technical writing at launch, opus 5 just set the bar so low that people ate it up anyways

This is compounded by a few specific style issues it has consistently, like overusing: - the words every/never as extraneous emphasis - numeric words (one/twice/third/etc.) which put focus on the number rather than the actual point - "X, so Y" sentences over section/paragraph-level structure

lukewarm take: opus 5.5 still sucked at technical writing at launch, opus 5 just set the bar so low that people ate it up anyways

real example sentence I've flagged: "A rule that names an identity regardless of credential type leaves the principal unscoped and matches on the attribute" like I as the SWE know what this is trying to say but even in context this is just a weird, hard-to-parse sentence structure

lukewarm take: opus 5.5 still sucked at technical writing at launch, opus 5 just set the bar so low that people ate it up anyways

a trend "<insert popular agent harness here> sucks, literally anything else is better" "do you have any suggestions with decent UX" "oh uh I made mine for myself so uh well, hm"

I'm pretty sure I've seen others come to this conclusion already periodically, but I think one consistent (but probably not inherent) weakness of LLM writing is that it fails to understand its readers and I wonder if this reflects a lack of understanding by the human user or an underspecified task

Anyways I need to wrap up a newsletter that is long overdue because I don't write quickly and agents keep giving me annoying output, so I probably won't try it anytime soon, either Ideas are meant to be free

I'm pretty sure I've seen others come to this conclusion already periodically, but I think one consistent (but probably not inherent) weakness of LLM writing is that it fails to understand its readers and I wonder if this reflects a lack of understanding by the human user or an underspecified task

And in the case of bad human writing (of which I keep far fewer samples), do we see the same issue there consistently where the author misunderstands the reader, or is it usually something else?

I'm pretty sure I've seen others come to this conclusion already periodically, but I think one consistent (but probably not inherent) weakness of LLM writing is that it fails to understand its readers and I wonder if this reflects a lack of understanding by the human user or an underspecified task

I never get good technical writing out of agents but I do find that the more I specify what the target reader does/doesn't want to know, their temperament, and what I do/don't care about conveying, the better the result gets, if only to a point (so far)

I'm pretty sure I've seen others come to this conclusion already periodically, but I think one consistent (but probably not inherent) weakness of LLM writing is that it fails to understand its readers and I wonder if this reflects a lack of understanding by the human user or an underspecified task

I think this only never occurred to me because of tool search and code execution, but MCP Resources are basically just getter tools you rarely use Any server-owned info for the agent that only (sometimes) has a useful meaning in the context of a prior tool result should maybe be a Resource

Though notably this all only holds true if (1) you don't also have a proper tool for the same info (otherwise the agent will almost always use that instead) and (2) your agent already has a native tool for retrieving MCP Resources (both CC and Codex do but I think a few popular harnesses don't)

I think this only never occurred to me because of tool search and code execution, but MCP Resources are basically just getter tools you rarely use Any server-owned info for the agent that only (sometimes) has a useful meaning in the context of a prior tool result should maybe be a Resource

this is sort of directionally true in the sense that labs know what they've optimized their own models for and have well-funded harnesses except that implies SOTA is following whatever the labs are prioritizing by virtue of not being you the labs are not necessarily optimizing for what you want

Still trying to formulate this thought correctly but I think a lot of the reason why new models feel less impactful in day-to-day work at Big Companies is because the SDLC puts constraints on how work is actually organized and how risks are managed that RL is inherently unsuited to

Adapting these systems to agents is difficult because these are all things that involve a certain degree of making judgement calls and hedging risk against authority, which are somewhat technical problems but are mostly complex social negotiations that agents consistently seem not to understand

Still trying to formulate this thought correctly but I think a lot of the reason why new models feel less impactful in day-to-day work at Big Companies is because the SDLC puts constraints on how work is actually organized and how risks are managed that RL is inherently unsuited to

I think a lot of organizational tensions around making full use of agents are a reflection of this, too

Still trying to formulate this thought correctly but I think a lot of the reason why new models feel less impactful in day-to-day work at Big Companies is because the SDLC puts constraints on how work is actually organized and how risks are managed that RL is inherently unsuited to