We're definitely in the era of the harness, with LLMs serving as logic circuits that can be swapped out. This paper effectively shows that the performance of the harness is the feedback loop. I'd expect this to be an increasing area of competition, especially with delightful dark horses like Hermes from @NousResearch appearing out of the blue.
https://t.co/HE4IHpVv1B
Are we moving away from CUDA lock-in? Either way, Qualcomm scooping up modular will give Qualcomm and thus developers a cross-vendor compiler/runtime for AI compute. Good to see.
https://t.co/lKjsmJ0yBc
@medcitynews The gap is workflow integration. AI pilots fail when they're slapped onto existing clinical workflows instead of embedded at the actual decision points. The fix is designing around patient signals and clinician decision moments, not adding another dashboard.
@emollick I see this daily as a forward-deployed engineer. The gap between what current models can do and what enterprises are actually using is staggering. The bottleneck is change management leading to integration, workflow redesign, and trust. The overhang is real and addressable today.
Absolutely love the new claude code /insights command. Thank you @AnthropicAI , @bcherny and claude code team. I've never been so in love with programming.
"You are a prolific, automation-heavy power user running Claude Code at an extraordinary scale โ nearly 11,000 sessions, over 9,100 commits, and 41,000+ messages across just seven weeks."
Observation from Davos this week:
The conversation shifted from "will AI take jobs" to "which jobs change how."
Two categories emerging:
Jobs that survive
- Require physical presence
- Involve regulated judgment (legal, medical decisions)
- Need relationship continuity
Jobs that thrive
- Orchestrate AI at scale
- Make decisions in ambiguous situations
- Create things that require taste
The interesting part: survival jobs might not pay more.
But thriving jobs require skills most people haven't developed.
Demis Hassabis told undergrads to "get proficient with these tools."
That's table stakes.
The real question is: proficient at what?
Using AI as autocomplete is survival.
Using AI as leverage is thriving.
Genuine question:
What does "high agency" actually look like in practice?
Everyone talks about it like it's a personality type, but I think it's more like a set of daily habits.
My working list:
- Defaulting to action when uncertain (vs. waiting for permission)
- Treating blockers as puzzles (vs. reasons to stop)
- Learning in public (vs. waiting until you're ready)
- Owning outcomes (vs. blaming circumstances)
But I'm curious what others would add.
What's on your "high agency" daily checklist?
I learned something building this week. I spent 3 days trying to get an agent to do something complex. Kept tweaking prompts, adding context, adjusting parameters.
Finally realized: the problem wasn't the agent.
The problem was I hadn't actually decided what I wanted.
"Make it better" isn't a spec.
"Reduce response time by 40% while maintaining accuracy above 95%" is.
AI doesn't expose the limits of technology; it exposes the limits of your thinking.
The best builders I know aren't prompt experts. They're clarity experts.
5 questions every company should be asking right now:
1. Which of our processes exist because of coordination costs, not value creation?
2. Where are we paying for expertise that AI has already commoditized?
3. What would our org chart look like if we redesigned it around what humans uniquely contribute?
4. Are we training people to orchestrate AI, or just to use it?
5. What's our plan when the juniors we're not hiring become our competitors?
Most companies are asking "how do we use ChatGPT?"
The right question is "how do we restructure around what AI makes possible?"
2026 prediction:
The AI community and blockchain community will finally stop talking past each other.
Right now:
- AI people see blockchain as speculation
- Blockchain people see AI as a buzzword to pump tokens
What's actually happening:
- $500B going into centralized AI infrastructure (Stargate)
- Decentralized alternatives emerging (Bittensor, Ritual)
- On-chain agents handling real transactions (Autonolas)
The convergence point: AI needs coordination infrastructure, and blockchain provides it.
When 1000 agents need to work together without central oversight, you can't use a Google spreadsheet.
The biggest opportunity of the decade is building the rails for human-agent coordination at scale.
Hot take on the SOLO CEO hype:
Most people don't actually want to run everything alone.
They want:
- Clear direction without micromanagement
- Meaningful work without politics
- A team that has their back without any BS
The solopreneur dream assumes you've already figured out:
- What problem to solve
- How to sell the solution
- How to handle isolation
That's a small slice of the population.
The bigger opportunity isn't making everyone a solo CEO, it's building firms that use AI to strip away bureaucracy while preserving what humans actually need: purpose, autonomy, and belonging.
One-person businesses are one answer, but a new kind of firm is another.
Uncomfortable truth: You can't prompt-engineer taste.
Taste comes from:
- Thousands of decisions about what to keep and what to cut
- Exposure to both excellence and mediocrity
- The courage to ship before you're certain
AI gives you infinite options.
Taste is knowing which 3% to ship.
The people who think they can skip this by being "prompt experts" are building castles on sand.
Your taste develops through iteration, not optimization.
Viewing blockchain as money infrastructure is like seeing the internet as a fax machine replacement.
The real blockchain use case is coordination infrastructure.
Blockchain will be used to move value between parties and organize work between humans and agents.
When you have thousands of AI agents working with hundreds of humans, you need:
- Verifiable outputs (did the agent actually do the work?)
- Trustless handoffs (agent A passes to agent B without central oversight)
- Programmable agreements (if X happens, then Y executes)
That's organizational infrastructure for the AI era.
The blockchain community keeps building exchanges when they should be building coordination layers.
I've been running 5+ AI agents daily for the past 3 months. Here's what I've learned:
1. Agents don't fail gracefully. They fail confidently.
You need good prompts, but you need checkpoints more.
2. Context is the bottleneck, not capability.
The same agent with better context outperforms a "smarter" agent every time.
3. Orchestration is judgment work.
Knowing which agent to deploy when is the skill.
4. Most agent failures are spec failures.
If you can't articulate what you want clearly, no agent stack will save you.
The limiting factor isn't the AI, it's your ability to think clearly about the problem.
Every AI tool that "does the work for you" is creating two classes of users:
- People who understand what they're automating, and
- People who just press buttons
The first group builds leverage.
The second group builds dependency.
A divide is being created based on how people choose to adopt these tools.
High agency isn't a personality trait.
It's a choice you make every time you use AI.
Genuine question:
If AI can do 80% of knowledge work, what's left for humans?
I keep automating tasks I used to think were "my job." I also keep automating tasks that I used to think were a lot of other specialist's jobs.
My current to-do list:
- Use judgement in ambiguous situations
- Maintain relationships that require trust
- Be tasteful in creative situations
- Carry accountability for major decisions
Is that enough for 8 billion people? Iโm not sure.
What is on your to-do list these days?
Drive beats experience every time now. That "architect + AI agent" setup is not 3-5 years away; companies are rolling it out right now. What does that mean for you?
Work isnโt not about junior vs. senior anymore, it's high-drive vs. low-drive.
A newbie who learns quick, ships code non-stop, and asks smart questions will crush it while a vet just churning out routine stuff and waiting for orders will be left behind.
This is great news for driven, high agency workers.
Solo businesses are trending right now, but do people really want total freedom, or do they want a solid structure minus the office politics? The solopreneur hype assumes you've got it all figured out: what to make, how to sell it, and handling the loneliness. Thatโs not easy though.
The real solution is to use AI to shortcut through red tape so your company can prioritize clear goals, work that matters, and a solid crew that's got your back.
One-person shops are cool, but they're just one angle on the future of work. Letโs also reimagine firms as launchpads for go-getters who thrive in a team environment
The shift is real, but the framing misses something important.
It's not about junior vs. seniorโit's about agency.
A junior who learns fast, ships constantly, and owns outcomes will outperform a senior doing rote orchestration.
The 70-80% reduction won't be evenly distributed.
It'll hit low-agency workers at every level.
The question for companies: how do you cultivate high-agency people, not just hire for existing seniority?
I'm personally a fan of the OPB and I operate a few at this point.
The advice is solid for high-agency people who already have the prerequisites:
- Know what problem to solve
- Can sell their solution
- Don't need external accountability
That's a small minority.
For everyone else, the bigger unlock might be firms that strip away bureaucracy while preserving structure. AI as organizational solvent could be an alternative model for lower agency folks.
The constraint isn't tools. It's knowing what to build.