Jesse Anglen

Founder @ 🚀 | Digital Labor Evangelist | AI Agent Pioneer | Speaker • Influencer • Thought Leader

[Jesse Anglen @Jesse_Anglen]

@AndrewYNg The more interesting pattern: how many people in this coalition haven't actually built with these tools. You can have sincere concerns and still have uninformed ones. Doesn't make them wrong, just limited.

[Jesse Anglen @Jesse_Anglen]

@svpino Skills as the unit of sharing makes sense if they're model-agnostic. But most aren't yet. You end up with Claude skills, GPT skills, Gemini skills, and you've just recreated the app store fragmentation problem at a different layer.

[Jesse Anglen @Jesse_Anglen]

@DrJimFan 'Vibe agents in the physical world' is a phrase that didn't exist 18 months ago. Wild.

[Jesse Anglen @Jesse_Anglen]

The inversion is what gets me. States have historically made citizens legible to themselves. AI flipping that, giving individuals the ability to make institutions legible, is genuinely new. Have you seen this actually change behavior when people do it, or is it mostly a transparency exercise?

[Jesse Anglen @Jesse_Anglen]

@GaryMarcus 4.6% sounds small until you run an agent that chains 20 decisions. Each step at 4.6% error compounds. By step 10 you're at roughly 63% chance of a clean run. That's not 'next to nil', that's a coin flip on anything non-trivial.

[Jesse Anglen @Jesse_Anglen]

@fchollet It shows up in how people build agents too. Everyone reaches for 'throw it at the LLM' because that's the only hammer they've got. Evolutionary algorithms, MCTS, constraint solvers, none of it exists in their toolkit. Same tunnel vision, different layer.

[Jesse Anglen @Jesse_Anglen]

@emollick The 'no impact' data gets distorted by averaging. We ran deployments in 2025 that moved real numbers. The common thread wasn't the tool, it was whether the team pointed AI at the thing that actually moves their business vs. the safest possible use case.

[Jesse Anglen @Jesse_Anglen]

Same crime shows up constantly in AI eval reporting. Benchmark points drawn from the same data distribution and time window aren't independent samples but they're analyzed as if they are. 'Consistently outperforms on 10 tasks' on correlated evals is this exact fallacy with different labels.

[Jesse Anglen @Jesse_Anglen]

The existing replies cover capex cycles and timeline differences well. The part that makes fiber truly incomparable: fiber didn't improve itself. You got exactly what you paid for. AI compute can potentially increase its own useful output over time. That's a structurally different return curve, not just a faster version of the same story.

[Jesse Anglen @Jesse_Anglen]

@drfeifei @cs231n @Stanford That department spread is a better signal than any benchmark. The tool crossed a domain boundary.

[Jesse Anglen @Jesse_Anglen]

Shipped my first vibe-coded feature to prod last year.

Worked perfectly. Tests passed. Nobody noticed.

Then a customer hit a weird edge case and I couldn't explain a single line of the code to my team.

That's the thing nobody mentions. Functionality isn't the finish line. Legibility is.

Karpathy officially called vibe coding dead this year and replaced it with Agentic Engineering. Human-supervised, auditable workflows.

Vibes don't pass security reviews.

[Jesse Anglen @Jesse_Anglen]

Spec-first works really well, I've been doing something similar. One risk worth flagging: a clean build against a wrong spec is worse than a messy build against the right one. First working version is a research prototype. Sometimes you discover the README was describing the wrong thing entirely.

[Jesse Anglen @Jesse_Anglen]

The 'model judgment' bottleneck is real, but it's specifically the self-correction loops that break. Small models don't catch their own errors well enough to course-correct mid-task. You can work around it with tighter process and human checkpoints, but it's not binary. Some agentic tasks have a higher judgment floor than others.

[Jesse Anglen @Jesse_Anglen]

The community self-selection is doing a lot of work there. Gist commenters usually found it by searching - they came for the content, not the feed. X rewards engagement velocity, so comments optimize for visibility. That's a completely different incentive structure, and different incentives produce different people.

[Jesse Anglen @Jesse_Anglen]

App Store productivity charts usually have Notion and Todoist type apps. A model runner being #8 is a meaningful signal. Consumer adoption territory, not just dev tinkering. Curious whether the downloads skew toward devs or whether regular users are actually running Gemma locally.

[Jesse Anglen @Jesse_Anglen]

The 'fit on a single page' part is the real kicker. It's not just fewer training examples. It's a completely different compression. A billion parameters to approximate what a handful of equations encode exactly. Have you seen hybrid symbolic/neural work actually closing that ratio, or is it still mostly open research?

[Jesse Anglen @Jesse_Anglen]

The research tracks, but the framing's missing something. Cognitive surrender is less about AI being dangerous and more about people not building enough intuition to know when to push back. Same thing happened with search and autocomplete. The fix is AI literacy, not skepticism as the default posture.

[Jesse Anglen @Jesse_Anglen]

@alliekmiller The read-only constraint is the real bottleneck. You can surface everything in context but actual triage requires action. The 'slow on purpose' part is interesting though. That's a deliberate safety call, not a technical limit. Worth watching when they flip that switch.

[Jesse Anglen @Jesse_Anglen]

@emollick The 39% less capital needed is the number I'd want to dig into. Is that because AI is replacing paid services, or because founders are making better decisions earlier and avoiding expensive mistakes? Both matter but they compound very differently.

[Jesse Anglen @Jesse_Anglen]

@svpino Same experience here. The weird thing is it's not just writing less code. The whole problem shifts. Less 'how do I implement X' and more 'what should even exist.' That's the better problem to be working on.