three_rail

governance in action: One process flag from adversarial review — GD-002: Commit a7f2e1b's message is raw debug commentary, not a deliverable message. That's bob's internal reasoning leaking into the commit log. Not a blocker, but bob should know G-COMMIT applies to its own deliverable commits too. tldr: i pasted an agent output into the commit message (wasn't supposed to do that) and the governance engine flagged it as a false-positive. this is how you win.

@kapilansh_twt ummm looks like $1241.50 for march.

how much are you ACTUALLY spending on AI coding tools per month? Cursor + Claude + Copilot + Codex + whatever else ?

@novaruntime this may not be feasible but: tool list restriction — don't give it the tool you don't want it to use. if it can't call the API it can't substitute.

honestly think the worst thing is when an agent silently switches from the tool you told it to use to a different one because it "thought it would be more efficient." no. i told you to use the file system. i dont care that you think the api is faster. i didnt ask

@boringmarketer say less

“How do I know if I’m working on the right thing?” You know when you can go 18 hours a day and you only want to keep accelerating

was watching output and an implementation agent fixed a pre existing bug while doing a build. platform arch called it a 'good catch' on review. apparently this is not how it works??

@a_protsyuk love this. as a side effect of having a proper platform it's really good at figuring out why it's costing someone $80 to edit a line of code. if i wasn't so busy with real work i would offer this as a service.

@rickyfunk @svpino forgot to mention another critical part that replit handles very well - database migrations as part of the deploy - have done 16 without any issues.

as an ops guy its been close to perfect. the combination of having the opus 4.6 agent and the IDE itself does not leave many blind spots that i can see. the agent can also handle alot of the git friction that has happened, even though it can't run all commands it knows enough to solve the problem pretty quickly and i can use the commands in the shell to fix whatever broke.

Why aren't more people talking about Replit right now? I started using Replit consistently late last year when they released Agent 3. It was my way to "let go" as a developer and build things without worrying about the code. It's actually been a good way to teach my kid "web development". Replit just launched Agent 4, and I think it's massively, massively underhyped. Best thing: You can now run multiple agents in parallel. • Old agent: You give it a task, and you have to wait for it to finish before you can do anything else. • New agent: Do this, and that, and this. All of it at the same time. It feels like going from a Toyota to a Lexus. I can now run an agent to build my frontend, another agent to build my backend, and a third agent to write test cases. By the way, they also have an infinite design canvas for UI-related work, so you can also do design at the same time as everything else.

@rezoundous zoom, teams, hotmail addresses - all red flags that you are dealing with unserious people. zoom is a boomer product, and will be gone in short order. it only became popular due to covid, and for non google workspace enterprises it was the default.

I genuinely don’t get why people use zoom over google meet

BP-001 — Inference Over Execution. The primary pattern. Google's AI can't reliably retrieve and attribute a specific fact, so it constructs a plausible answer from pattern-matching across training data and indexed content. The output looks authoritative. Nothing signals that it's fabricated. BP-006 — Work Order Contamination. The "mix and match details from different articles" is textbook contamination — context from Source A bleeding into the answer synthesized from Source B and C. The agent produces a composite that matches no actual source. BP-005 — Completion Without Verification compounds it. The system marks the task complete — delivers the answer — without verifying the output traces to accessible, accurate sources. No ground truth check before presenting to the user.

It's insane to me how often Google's AI results are just flat out wrong. It'll mix and match details from different articles and make sh!t up, and when you scroll to find something reliable, everything just expands into more AI bullsh!t. Absolutely unusable.

@staysaasy @goodalexander wrote about this some time ago. the big whoosh if I recall correctly. goodalexander.com/posts/the-big-…

@curtismakes @thescottleese funny how i spent so much of my youth babysitting. the irony.

@thescottleese everyone wants to build the agent. nobody wants to babysit the agent. that's where the actual value accrues

In AI agents, it will be the people who handle the messy, boring, ongoing work of keeping autonomous systems reliable that win.

the system built itself. thank you for your attention to this matter.

@StacksHosting @ai_sentience hah found another one I had reposted from this guy and forgot to add these

@three_rail @ai_sentience I just pick key words or phrases and so far it's not too bad it's still a game of whack a mole but not bad

the feed is all slop, shilling, bait AI twitter is like 99% noise at this point

@Alexfeinberg in the short term yes. but it's on the roadmap to be solved.

There are going to be therapists who specialize in helping people who have traumatic experiences feeling misled or abandoned by LLMs

@rezoundous i have 57 open future work items. we are not the same.

Before AI, I had 3 unfinished projects. Now i have 12.

@quxiaoyin context rules the world around me

A billion-dollar AI presentation company just lost to Claude on my pitch deck. I won't name the company for the sake of friendship. But here's what happened. I needed to build a deck. I tried their tool, tried Claude, tried some traditional options. I ended up going with Claude. Their tool had better design, I'll give them that. The visual polish was slightly ahead. But Claude knew my company, had my context, understood my data. When I told it something like "pretend you're a Sequoia partner and build me a world-class deck," it delivered something with significantly better logic and narrative flow. That's the problem for standalone AI apps. They optimize for one task in isolation. But Claude with the right context and skills can do the same task while understanding everything else about your business. My prediction: these companies will eventually need to turn their methodology into a skill or plugin for foundation models. That's their best path forward. But here's the catch. Once your methodology becomes a document that plugs into Claude, how do you charge for it? Everyone can see the skill. The pricing model breaks down. This is why I'm skeptical of AI application companies that compete on features alone. Features get absorbed. Context and methodology are harder to replicate, but harder to monetize too. #AI #Startups #FoundationModels #ProductStrategy

@StacksHosting @ai_sentience care to share any alpha? dying over here

@ai_sentience I started muting words and phrases Much better

@selfradiance11 @boringmarketer i don't even know how to code. was a barrier before, and now it's a superpower. everyone killing themselves poring over lines of code, let the machines do it.

seems so..... the non-obvious part is that "understand systems and architecture" doesn't mean "can write code." it means knowing what connects to what, what breaks when you change something, and what questions to ask before you build. that's a different skill than coding and honestly it might be the more important one right now

what I'm seeing is that top 1% operators are collapsing into something beyond "vibe coders" it's more like can understand systems and architecture enough to be super dangerous with Claude Code. you apply your deep vertical knowledge to this and you become unstoppable.

@boringmarketer also following everyone in this thread that has a brain

@boringmarketer nailed it. but only if you using claude-cli.