QWE AI Academy Why isn’t smarter AI actually moving the needle on your projects?
You’ve got access to models that crush benchmarks. You prompt them for strategy, code, analysis, even persuasion. The outputs look brilliant. Yet the project still stalls at the same places: waiting on data, approvals, other people, distribution, or messy real-world execution. Sound familiar?
That’s the point in Ruxandra Teslo’s essay “Intelligence is not the main bottleneck” (published Jul 21, 2026). It jumped on HN fast (123 points, 116 comments in the first stretch) plus the usual X and LinkedIn pile-on. Core claim: no matter how sharp AI gets, intelligence is often not the binding constraint on real-world change.
This isn’t another summary. It’s the hands-on version – audit your own work, then aim ChatGPT or Claude at the actual bottlenecks instead of minting more clever plans.
What “intelligence is not the main bottleneck” actually means for AI users
Clinical trials still chew through something like seven years and more than a billion dollars per drug. That number sits in Teslo’s piece for a reason. Pre-clinical science and molecule design keep improving; AI-biotech stories raise huge rounds – Chai Discovery, per the same essay, sat at a $3.8B valuation on designing better compounds given a target. And still Eroom’s Law holds: inflation-adjusted cost of a new drug has roughly doubled every nine years. Human data stays irreplaceable. Governance, locked datasets, incentives, and patents decide what gets tried more than raw smarts do.
Same drag shows up outside medicine. We’ve known how to build denser housing for decades. A lot of customer-service bots still feel worse than a tired human. GDP did not jump. Entry-level roles mostly stayed put. Tyler Cowen has been saying for a while that capabilities land faster than institutions absorb them. Ethan Mollick’s jagged frontier makes the mess stickier – superhuman on some hard tasks, oddly weak on easy-looking ones – so the bottleneck just moves.
For your projects that means: a smarter model rarely fixes “I can’t get the dataset,” “legal won’t sign,” “the team won’t change the process,” or “nobody will pay for it.”
Think of it like hiring a genius who can redesign your factory floor in an afternoon while the building permit still takes eighteen months. The genius is real. The permit is still the schedule.
Practical setup: 20-minute bottleneck audit for any AI project
Open a fresh chat. Paste this (swap in your project):
Project: [describe your current stalled AI-assisted project in 3-5 sentences - goal, what you've already generated with AI, where it is stuck].
List every step from idea to real-world outcome. For each step mark:
1. Is the main constraint intelligence/knowledge/creativity? (yes/no)
2. If no, what is it? (data access, human approval, coordination, physical execution, incentives/money, distribution/adoption, regulation/compliance, trust/reputation, time, other)
3. How long has this step blocked progress?
4. What would unblock it without needing a smarter model?
Rank the top 3 non-intelligence bottlenecks by impact. Suggest one concrete next action for each that an AI tool could help with (drafting, analysis, simulation, outreach template, process map, etc.). Be blunt.Run it. Most of the remaining work usually sits outside pure intelligence. That’s your map.
Common categories that surface:
- Data/access: locked datasets, missing human feedback loops, siloed internal docs – including multi-month waits (Teslo notes firms waiting a year for NIH imaging sets)
- Process/governance: approvals, compliance checklists, IRB-style reviews, change management
- Coordination: handoffs between people/tools, mismatched incentives
- Diffusion: packaging, trust-building, distribution channels, habit change
- Physical/real-world: manufacturing, logistics, patient recruitment, site activation
Save the ranked list. New priority queue.
Pro tip: Re-run the audit every two weeks. Bottlenecks migrate. Yesterday’s data problem becomes today’s adoption problem once the model improves.
Advanced usage: point AI at the real constraints
Once you know the bottleneck, stop asking for “better strategy.” Point the model at that specific friction.
Data access gap – Ask for public proxies, a synthetic-data plan that still flags where real validation is mandatory, or exact request/FOIA language plus an escalation path. Or shorter: “Rewrite this analysis so it only needs the fields I already have.”
Approval / compliance – “Draft the shortest one-pager for [regulator/legal/manager] that answers their top three historical objections. Include a risk matrix and a phased pilot with zero new liability on day one.” Second pass: role-play the skeptic and punch holes in your own draft.
Coordination / incentives – Map who actually decides and what each person loses if the project works. Generate talking points per person, or a shared dashboard mock that makes the win visible. Meeting sim: “You are the VP of Ops who hates new tools. Roast this proposal. List the three concessions that would get you to green-light a 30-day pilot.”
Diffusion / adoption – Skip another feature list. Generate the onboarding sequence, the before/after metrics users care about, an internal champion kit, and a short email drip that handles “this will never work here.”
What if the next model release changes nothing about your shipping speed? That’s the question the audit is built to answer before you spend another week on clever prompts.
For teams, dump the audit into a shared Notion or Airtable: bottleneck type, owner, AI-assisted artifact, human action required, status. Weekly review. AI drafts and analyzes; humans own relationships and final calls. If you already process-mine workflows or keep tight human-in-the-loop eval sets, fold those artifacts into the same board – they hit the non-intelligence layer directly.
Honest limitations of this framing
Intelligence is not worthless. Better models still widen the tractable slice – molecule optimization, code gen, first drafts, pattern spotting. Jaggedness means some bottlenecks will melt as capabilities move. Pure software or research-only work can still be intelligence-bound.
The cope risk is real. “Bottlenecks elsewhere” becomes an excuse not to ship and not to learn the tools. Some people will drown in process theater and under-use what’s already available. Exact attribution stays fuzzy too: no clean public study splits biopharma slowdown into “X% regulation vs science vs incentives.” Treat the essay as a directional diagnostic.
One structural trap the essay flags: patents reward novel chemical matter (composition of matter) more than novel biology or target validation. Money and talent herd into molecule design – the Chai-style, given-a-target problem – while harder target work stays under-incentivized. More intelligence on the easy part does not automatically expand the frontier.
HN threads also keep repeating a blunter point: persuasion benchmarks can beat human debaters and still change nothing. Reputation, skin in the game, and institutional authority block the path. Hyper-persuasive text alone rarely moves a regulator or a committee.
FAQ
Does this mean I should stop chasing better models or AGI progress?
No. Use the best tools you can get. Just stop assuming the next capability jump auto-unblocks your real goals. Aim the smarter models at the friction the audit surfaces.
How do I apply this if I’m a solo developer or writer, not in biotech?
Same audit. Usual bottlenecks: distribution (eyes on the work), trust (why switch), consistency (shipping cadence), feedback loops (real user data). Point AI at distribution assets, objection handlers, and a lightweight measurement plan – not endless feature ideas. Concrete: a writer stuck on growth asks for three platform-specific hooks plus a simple tracking sheet, not another outline.
Isn’t this just “execution matters more than ideas” rebranded?
Partly, but sharper. Classic advice still treats intelligence and ideas as the scarce input. Here the claim is stronger: in a lot of domains intelligence is already abundant (or about to be), so the marginal return on more of it is low while returns on unblocking data, process, incentives, and diffusion stay high. The medicine timelines and the slow economic absorption make that distinction concrete. You still need good ideas. You just can’t stop at the idea.
Open your current stalled project. Run the audit prompt. Pick the #1 non-intelligence bottleneck and generate one concrete artifact that attacks it today. That’s the move.