Anselme

Building the LLM-powered assembly line for factual question answering.

Builder. @NotreDame

[Anselme @theMissionBoy]

@ciengalon @sflorimm it was clear by late Feb that gpt-5.3-codex was a more reliable coding agent than opus.

the type of shortcuts opus 4.5/4.6 took were just infuriating. haven't touched it since then.

[Anselme @theMissionBoy]

every once in a while you gotta remind yourself that you're not a car

[Anselme @theMissionBoy]

@benhylak the OP was completely right

pretty much all those things happened

[Anselme @theMissionBoy]

whatever it is that you are doing, not the day to day stuff but the multi-decade north star, never forget why you chose that path because everything else is secondary

[Anselme @theMissionBoy]

codex might be the most slept-on product of all time

[Anselme @theMissionBoy]

@astridwilde1 "Tests should be about quality, not quantity"

in AGENTS.md or CLAUDE.md

[Anselme @theMissionBoy]

planning to start sharing long-form articles of what i am learning. draft now; artcle coming soon.

i used to be obsessed with TDD before coding agents forced everyone to start writing tests

recently discovered an effective way to TDD model-facing code (non-deterministic etc).

[Anselme @theMissionBoy]

@benhylak The combination of order-of-magnitude speed increases and GPT Live will unlock lots of use cases we can't even imagine today

[Anselme @theMissionBoy]

commodity software, the type you used to get from dev shops or UpWork, is a solved problem

[Anselme @theMissionBoy]

@thdxr It's difference between a dirty house and an unstable one

[Anselme @theMissionBoy]

@alighodsi Is there a standard yet for defining ontology or semantic authority for agentic workflows?

[Anselme @theMissionBoy]

Trying to make AIs return the same answers, when nothing has changed.

[Anselme @theMissionBoy]

@MikeMcCormick_ Verification as in correctness or permission or both?

Working on the former for factual accuracy.

[Anselme @theMissionBoy]

Sol seems proactive in a way I haven't seen with the prior models.

When there is enough information to change course, it's fast and doesn't waste time.

Judgement is not perfect but baby steps

[Anselme @theMissionBoy]

somehow surprisingly, 'medium' reasoning effort seems to be much better for long running tasks than high or extra high.

[Anselme @theMissionBoy]

the biggest difference between gpt 5.5 and gpt 5.6 sol seems to be that the latter is more strategic and really good and creating tools and infra for achieving the final goals

[Anselme @theMissionBoy]

learned most of these lessons at @8thLightInc

Quality software has never been more valuable

[almonk @almonk]

https://t.co/Rz0q1W33ut

[Anselme @theMissionBoy]

Working on a project that takes an input through 8 model calls, each taking the output of the previous (after some transformations)

Accidentally landed on one-model-call "harness" that allows to run experiments on one step and getting stable results before integrating into the whole pipeline

Coding agents are so good with this setup that the quality and pace of work became ~3X all of a sudden

Not sure if it applies to all projects but it's a big unlock where applicable

[Anselme @theMissionBoy]

@isidoremiller "many problems can be reworked to fit this!"

a lowkey big insight

[Anselme @theMissionBoy]

@mattzcarey @RhysSullivan @zeeg maybe a clearer framing is

"with MCP, as a business, you don't own or see the full 'question to answer ' user journey. In essence, you have been disintermediated from your own users/customers."

that's the part I'm unsure everyone understands and the full implications