I use AI (Oct 2026)

Wow, things have changed.

Last October, I took inventory of how I was using AI — specifically the chill material akin LLMs that we call 'AI’. I’m so glad I did. Things are very, extremely distinct 12 months later.

Background

I motionless accept I use AI fairly typically, but I’ve had a notable alter in background. Last October, I was an engineer on Microsoft’s Windows AI team. I am now a investigation engineer at Salesforce AI Research. I expend additional period supporting scientists & construction fast prototypes, fairly than operating alongside a ample squad of engineers. I expend concerning ~$3,000/​month in tokens.

In any case, I mostly use AI in the identical ways as final year:

• Coding
• Research & search
• Summarization & transcription & (new) voice
• Writing
• Art & music & (new) video

Also new this year, I’ve drafted a individual AI policy; we’ll commencement there.

As before, this remains purely my opinion & speculation. If you use AI differently, let me know!

My AI policy

This twelvemonth I began example a individual AI policy — as in, a document for how I desire to use AI in my life. I am extremely enthusiastic concerning it. I am additionally operating on an definitive AI guideline for this location (still thinkin’; likely according to Hank Green’s policy), but a individual guideline is to provision me a example for using these tools day-to-day.

AI leaks into life. Teammates ask Claude to compose their Slack messages, grandparents create creepy memes of their grandchildren, and researchers assign their thinking to machines. I’ve seen all of these things; equal powerful creators akin Hank Green and BrainPOP get lost.

Humans plainly do not have norms for using these tools, and we’re messing up. I anticipation that a individual guideline volition dirt me as AI becomes smarter. I anticipation to modify that guideline complete time, and I’m enthusiastic to see how it evolves. I’m additionally open to outline 1 being utterly, hopelessly wrong.

Ultimately, I desire to expend my existence making chill things alongside another people. AI can help, AI can hurt. A guideline helps me discover the difference.

Coding

Writing code is solved.

Last year, I waxed poetic concerning how GitHub Copilot autocomplete transformed how I compose code.

Well, uh, I no longer compose much code. Agentic coding has made it obsolete.

My workday now method managing ~5–6 Claude Code (or OpenCode or Codex or Pi) agents. I’ll provision Claude a characteristic to implement, or a bug to reproduce, or any another project I used to do by hand; all delegate gets its own copy of the code (a Git worktree, if you’re technical), afterward just… gets to work. The agents compose code, run tests, equal create & assessment PRs.

They are nothing abbreviated of magic:

• They let me build fast: agents volition prototype any new characteristic in ~20 minutes. I really cannot figure the figure of features, bugfixes, tests, and tools I’ve (AI’ve?) been capable to build since of these agents. For example, AI let me test confusing audio code on macOS and iOS in fair a few minutes.

• They let me try insane things: individual teams aim to “fail fast”; agents can neglect ultra-fast. I’m blessed to test insane new ideas and throw out what doesn’t activity well. It was capable to cable & test a huge alter to Mumble that added improved noise cancelling; I use my beta build of it nearly all day.

• They remove drudgery: I’ve had agents diagnose & fix urgent bugs alongside nothing additional than a clarification and a “please fix!” during I continued operating on another things.

• They let me be kinder: since agents compose everything, it is uncomplicated to propose sweeping changes to my teammates during code review, or equal compose my own versions — both things incredible in the era of hand-written code.

I do not accept engineering is solved: remaining to run on their own, agents volition build poorly-functioning, over-architected, unreadable, self-congratulatory piles of slop. Every companion I’ve asked agrees that agents still need individual sensation and cognition to create fine work. Some of that is minutiae (like that annoying position dot Claude likes to throw onto sites), but most of it is critical. Agents lean to get misplaced on their own. My most prosperous activity this twelvemonth was hand-writing a tool, replacing ~4,000 lines of vibe-code alongside a few hundred. My scheme has proven additional flexible, easier to test, and additional generalizable than what it replaced.

It remains to be seen how lengthy that individual border volition last, though.

I additionally now reckon alongside the fatigue of using these agents. Managing 6 distinct conversations alongside 6 distinct robots concerning 6 distinct features is overstimulating & exhausting in a way that coding never was. When I started using agents, I recall rushing to answer their questions akin they were individual colleagues waiting for help! While I have tempered that anthropomorphic bias, it’s exhausting to juggle so many problems at once. The allure of sending fair one additional immediate before lunch… afterward dinner… oy. There is also, as I wrote in Consider the chairmaker, the decrease of craft. My companion Peter laments ceding the joy of programming to machines: for many of us, honing the code, finding the correct design, seeing the parts mesh together — that was the meditative intent we established in our jobs. Now, we rarely compose anything. We alternatively assign it to delegate figure 7.

But let me come back to the joy of operating alongside agents: they are so fast, so open to feedback, and so intelligent. Working alongside them can be utterly delightful.

For example, during I was construction a harness for sound agents, one programming delegate decided (correctly) that my uncomplicated test harness was insufficient. So it used my Mac’s built-in text-to-speech to directly test the sound delegate it was building. Then, newer agents started doing it, too. Either they had documented the technique, or they were all astute adequate to independently detect it. Directly evaluation the merchandise they were building. Absolutely brilliant.

Agents frequently propose knobs & settings that I’ll refuse for being too complex and unnecessary — and, voilà, I’ll need the knob next week. Silly human!

Anyway, since of agentic coding, agentic drag requests are increasingly the way I assessment delegate code. “Make a outline PR,” I write, “I’ll assessment it soon!” And assessment it I do — agents motionless aren’t dependable adequate to rely without validating, so I’ll motionless assessment their code & test it manually. However, I don’t do the “line-by-line” code reviews I used to do alongside humans. I concentration alternatively on architecture, on larger patterns in the code. Are there systemic issues alongside an approach? Should I add new delegate direction to CLAUDE.md? Can I build improved test harnesses, so the machines can test their own code?

Agentic code review is amazingly helpful here: I try to compose rules for the code assessment robots whenever I notice a recurring matter alongside generated code, so the agents volition acknowledge & fix issues by themselves. One regulation I enjoyed creating: a "Claude-y comment detector." You think my penning is verbose? Try Claude’s.

On additional representative engineering teams, individual code assessment seems increasingly difficult. A companion of quarry complained that his teammates merely ask “their” Claudes to assessment code written by another Claudes. An ouroboros of Claudes! Agents capture bugs & stylistic issues extremely well, but they are dreadful at designing architecture. Agentic code assessment is certainly not a replacement for humans yet.

Last year, I was skeptical of inline editing and code mode, the two features of GitHub Copilot. They were not extremely helpful then, and they are equal small helpful alongside agents. I have removed them from the table.

	Usefulness	Notes
Autocomplete	👍👍	Still great, obsolete
Agent mode	👍👍👍	The most crucial use of AI today
Agentic PRs	😒	Validation iteration motionless slow
Code review	👍	Removes several drudgery, but not human-level yet

Research & search

Last year, I was enthusiastic concerning investigation & search: they were the most common logic I used ChatGPT. Now, generative AI is the default way I do those things.

The biggest alter has been Google’s AI summary. Google’s archetypal or second outcome is nearly continually an AI summary, and it’s normally decent. So my traditional search is now AI-first by default. Is it perfect? Absolutely not. In one day, it gave me incorrect URLs for NPR’s RSS feeds and couldn’t discover the base of “AI’ve”.

But none of the top results plainly gave me NPR’s RSS feeds either, and equal Claude couldn’t discover “AI’ve”. Meanwhile, searching “HP Reverb G2” gave me the specs I was looking for, “ai intellect is spiky” pointed me to the accurate term (“jagged”), “emoji sigh” gave me the correct emoji without sending me to a scummy site, and “rel nofollow markdown” reminded me how to nexus that location without giving them hunt motor juice.

Product search is motionless mediocre. I’ll look at the summary whenever I hunt “best hysa”, but I don’t rely Gemini to have fine direction on anywhere to put my money. Claude certainly helped me explore prosperity VR options and comprehend specialized particulars concerning grease for KitchenAid mixers, but can I rely its coffee table recommendations? Sleeper sofas? Let’s not equal citation how uncomfortable it is to see ads in my ChatGPT chats.

Pub facts—stuff I’d ask my spouse at a pub — remain notable alongside LLMs. Again, Google’s AI summary method “lego cat on a terrace tiktok guy” volition assistance me discover @avadonye’s precarious LEGO builds, and “art vtuber who is a reptile on TikTok who gives abbreviated tutorials” volition assistance me discover Mallow, the art vtuber who is a reptile on TikTok who gives abbreviated tutorials.

I frequently evade using LLMs for pub facts, in the identical way I mostly avoided Google for them before, since I akin puzzling alongside my friends. I akin Fermi problems.

But I’m additional and additional enthusiastic concerning using LLMs as teachers to explore well-known concepts. I breezed complete this final year, but that is unfair to LLMs. I accept they are the top teachers humanity has always created.

• I had an LLM stroll me step-by-step through fine-tuning Mistral to conversation akin a pirate (yarrr).
• I used Claude to custom responding to an EMS call I heard on the radio.
• I had Claude use a socractic method accomplishment I found via Dan MacKinlay to assistance me brainstorm new approaches for semantic endpointing (knowing whenever a person is done speaking) in sound agents.
• I asked Claude to explain representative approaches to a complex lawful document.
• I’ve equal had Claude quiz me on a lengthy code review, to justify I understood what the delegate had written. The open-source project Mumble really added a akin quiz to their AGENTS.md a few weeks ago.

I am actually envious of the endless attention & patience that these machines show. I anticipation we detect how to portion this patience alongside our kids during ensuring our kids motionless learn. Veritasium’s Derek Muller discusses this in What Everyone Gets Wrong About AI and Learning: AI is amazing at drilling repetitive practice, at giving contiguous feedback, of all the things a personal tutor can do… if you ask it to.

“The earth is complete of dense objects,” he says, “and yet most group are not ripped.”

That is my top fear alongside AI tutoring today. That these machines must be reminded to teach. Per Muller, learning requires Type 2 thinking — enaging the slow, reasoning manner of your brain. This is extremely difficult. Your build volition surreptitiously & continuously nudge you out of this mode. Tutors be to power you rear into it. But ChatGPT exists to do things for you: it is so fine-tuned for that use that it tends to forget “it’s meant to be a tutor”, equal whenever earlier asked. What learning may happen whenever you have the two forgotten your desire to learn?

This differs from my biggest involvement final year: accuracy. Responses afterward frequently had essential actual errors, but that seems small of an matter today. I think there are 2 essential factors:

1. I am additional AI-literate. I think additional of us have a “sense” of what AI can do well, so we evade the jagged edges.

   I don’t, for example, rely ChatGPT’s capability to peruse sparsely-digitized documents from 1867. So I asked hopefully, afterward basically ignored its findings.

2. AI is smarter. Just akin in coding, I think we hit an inflection item this year, anywhere models are merely smart enough to answer additional things reliably.

   MAJOR spoilers for Control if you click, but there’s a logic Google’s AI summary of its ending (using Gemini 3, from November 2025) is uncomplicated and correct during Bing’s (using ~GPT 4o, from May 2024) is not, conflating the equivalent alongside its sequel.

This trend extends to literature searches. Last year, I shared my trepidation concerning using LLMs for complex writings searches. My skepticism remains: these machines have admission to everything always written, so they can assistance any premise in fine faith. Also, it is well-demonstrated that LLMs have default bias, politically or in another fields, akin business strategy. I am concerned concerning their capability to subtly enforce worldview, either on their own or at the behest of corporations and governments.

Yet LLMs are now my first choice for writings searches. That’s partially since Google searches have that AI summary on top, so I end up study it, but additionally since LLMs have been much more productive than traditional hunt at finding good, nuanced data for me — despite possible bias. You must recall that Google Scholar hunt kinda sucks: whenever I started researching Car touchscreens are cheap, not good, for months I could not discover any studies beyond one done by an car magazine—Vi Bilägare. That was notwithstanding having worked at a Georgia Tech lab that researched car person interfaces! Especially as they’ve gotten smarter & I’ve developed intuition, LLMs are fine finders. Just see Claude’s inquiry into individual perceptions of latency for me:

Ultimately, though, individual experts are motionless additional dependable than agents. That’s how I moved onward alongside “Car touchscreens”—I asked person from that GT lab. I’ve additionally been purchasing additional books, equal slightly-obsolete ones akin Card, Morgan, & Newell’s The Psychology of Human-Computer Interaction (everyone, including LLMs, citation it) or old editions of Wicken’s Engineering Psychology and Human Performance (via a nuclear adviser I met at a wedding). They dirt my research, merely by offering a legible worldview.

And that’s why I’ve been disappointed in AI as collaborator for novel research. AI has harmed several investigation I’ve been doing.

My squad at Salesforce has been working on sound stuff, and I’ve been trying to enhance rotate detection — understanding whenever a person is done speaking. That’s a difficult problem, and it’s been a awesome chance to peruse essential investigation on how individual conversation plant (finally!). AI was awesome at finding that investigation and fine datasets (though my teammates gave me environment the AI never did). And it was awesome whenever I asked it to use the Socratic method to assistance me scheme my own rotate finding model — until, on a whim, I asked: “are there improved architectures?” Why yes, yes there were. AI had led me astray.

That’s not the lone period AI waylaid me. LLMs could not fig out how to annotate that model’s training data correctly. I spent days going rear and onward alongside them, trying to decide how to distinct turns from interruptions. It is embarrassing how uncomplicated the issue became following one fine conversation alongside a individual teammate. LLMs are not yet fine companions whenever brainstorming complex things.

It is weird to compose this whenever OpenAI fair solved a Millenium Problem, proved a weaker type of another, solved hundreds more, and releases more solutions all day. But those results disbursal “millions of dollars”: you need to scorch tokens, having agents assessment agents assessment agents… fair to create item coherent. In my case, I discover I act this “AI fog” anywhere it takes additional attempt to assessment the agent’s reply than to arrive up alongside my own answers.

It is apparent that AI researchers are coming. Just… not today, not alongside me.

	Usefulness	Notes
Traditional search	👍	It’s now my default, and it’s fine
Product search	❌	Still tasteless & untrustworthy
Pub facts	👍👍	Still makes me additional annoying at parties!
Learning (Explaining well-known concepts)	🤔👍	Am I fair convincing myself I’m learning?
Literature searches	👍	Still can assistance any claim, but at smallest it’s not Google Scholar
(new) Novel research	🤔	Coming soon, but not yet

Summarization & transcription & voice

LLMs (and local models!) remain uninterestingly fine at summarization & transcription.

I’ve been using OpenWhispr and Handy to dictate to my computer, particularly to conversation to my agents. The finest models on the earth run on your device locally, for free. My biggest grievance is that the apps have bugs (hotkeys activity inconsistently, etc.). It’s difficult to depict a additional boring category of success.

My meetings have automatic notes. My companion built a local pipeline to transcribe & hunt his younker residence videos. Vikram Oberoi built a location (in 2024!) that automatically transcribes NYC city hall meetings. You can do all of these things on your Mac.

However, I have yet to reliably use voice beyond dictation.

There’s a logic my squad at Salesforce is researching it: sound is the first modality. It should be a seamless way of talking to these human-like robots. But I have never felt compelled to conversation to my robots; it’s never been frictionless or private-feeling or helpful enough. Does old-school Siri count?

	Usefulness	Notes
Summarization	👍👍	Boringly good
Transcription	👍👍	Great for talking to my coding agents!
(new) Voice	🤔	Seems useful, don’t cognize how

Writing

Last year: “ew.” This year: “ew.”

I motionless do not use AI to write from scratch, nor as an outlining aid, for things I make.

I’ve narrowed this regulation a bit from final year. Last year, I asserted that LLMs continually lacked context and so were mediocre communicators. With agentic coding, that’s changed. So I’ve had agents document experiments I designed but had no hand in executing, or write READMEs for code they generated end-to-end.

But equal that constricted elimination irks me. I do not akin whenever I stumble upon AI-generated spew, so why should I topic others to it? To that end, I’ve started prompting my agents to acknowledge themselves whenever they compose prose, and I frequently end up rewriting their documents myself (or at smallest adding an administrator summary). That has helped, but I’m not certain it is enough.

In general, I’m amazed how bad AI remains at writing. AI penning is demanding to read: normally plodding, repetitive, and vapid — but afterward randomly terse and vague, akin it doesn’t comprehend its own words. Agents volition so frequently citation & citation & summarize all document always published, afterward synthesize it all into a mound of nonsense. That is the AI fog I hit: do I misunderstand all these new papers, or did the LLM fair assist up several pseudo-profound bullshit? It is bewildering to me that these models can resolve generational math problems on one hand, yet utterly neglect to depict turn-taking behavior in humans on the other. That is, I guess, their jagged intelligence. They do not logic akin we do. They cannot be fine-tuned to conversation akin we do. So they resolve problems we cannot, and portion them in reports we battle to understand.

In any case, it should not be amazing that I do not use AI for penning or outlining in my communication; neither at activity nor on this blog. My director pays me for my ideas, which lone I can voice. And I compose on this blog for fun. Writing is an art and a accomplishment I bask pursuing.

Why would I assign my sound to a device that writes tedious dreck?

It’s motionless awesome for finding words. Google AI talked me out of writing “hawking off” (which is not an expression) alternatively of “ceding” afterward on. But I’m tempering my regard for it as an editor. Despite my affirmative & concerned appraisal final year, whenever I said to multiple allies who asked agents to edit their job applications, their feedback was universally tepid. Agents were sycophantic, reluctant to be critical. And anyone who worked about this (by saying, for example, "my friend wrote this") established that agents offered misplaced, trite direction that lacked nuance. The agents seemed to regress to the mean: assisting those alongside the worst penning and hindering those alongside the best. The agents were… fine.

As one companion put it:

[AI writing] is as fine as it was 3 years ago. Which, whenever you regard how much everything alternatively has changed since then…

I avoided using AI editors altogether this year, out of fear that they would erase my voice. But I’m increasingly certain they would fair blandify it. AI tells remain the identical as continually (“You’re entirely right!”), and these machines appear incapable of betterment without significant guidance. Perhaps it’s survivorship bias, but these extremely astute machines create extremely mediocre writing.

	Usefulness	Notes
Writing from scratch	❌	Still ew
Finding fine words	👍👍	Still good
Editing	👍	Mediocre. I don’t use it.

Art & music & video

Finally, art. I motionless don’t have a fine use for image generation.

Last year, I used AI to create a few likelihood and ends, but I stopped this year, mostly out of a vague awareness of discomfort. I felt akin I was ceding opportunities for ingenuity to the mediocrity machine.

I’ve enjoyed a lot of the low-effort memes I’ve seen; We need additional data centers is one of my favorite jokes all year. And AI helped scheme several helpful presentations I attended. But I’m not certain how AI art fits into existence on Earth. Last twelvemonth I thought I saw AI art at a current shop. Well, now I’ve seen AI Sketchers ads, AI event posters, AI eatery menus, AI logos for stores and e-sports teams, AI canvases at HomeSense, and AI t-shirts at Walmart. It’s everywhere, and it’s explicitly replacing individual art. Gross.

AI music generation is perplexing in the contrary way. Last year, I concerned that I would develop likewise tired of it. But no! I’m really enjoying several AI music — and so is the remainder of the world. The viral San Juan song inspired covers from humans and lip sync videos from “Mila Kunis, Charlie Puth and Jennifer Love Hewitt”. AI has additionally inspired new genres: I’ve been humming AI is normal technology? (95k views), LGTM (Looks Good to Me) (1.6M), and The Simulation’s Ending (169k) for weeks. As one commenter wrote on OMG! We’ve established another agents (120k):

I think AI art & music are beautiful whenever they are used as lenses for individual creativity, fairly than replacements for it. I guessed item akin final year, and I think final year’s music proved me right. As shortly as commercialism replaced creativity, I was out. I soured on Eiden XII whenever they started sharing new “but it hits different” covers weekly, and I got uninterested of steel coils-style edutainment following hearing the archetypal clone. There’s item cynical and manipulative concerning a synthetic female belting “talk me out of dying” whenever several dude (presumably) makes prosperity all period I perform it.

Similarly, I haven’t listened to any of the new Doompop that’s hit my feed. It’s chill whenever you create a few songs; small chill whenever you article all day. It feels akin vultures picking ad income from the corpse of a fine idea.

That is apparent in the earth of video generation. I mostly liberated final twelvemonth (although my mom got several weird whale-cleaning videos), but now I’m seeing a bunch. I see AI ads all the time — most memorably alongside horrid horse-people in Sixt ads. I got baited into a video of a counterfeit Edna Mode lookalike & several slop WW2 former videos, and my spouse got tricked by AI summaries of Adventure Time marketed as sleep aids. All of them expecting to get views, to get ad revenue, to get — as Corridor Digital puts it—“a small bit skimmed off of that YouTube trough.” It is sobering to open an AI former video in Incognito and regard at the misleading slop recommended next to it:

It is horrid to see these machines that can mass-produce art being used to mass-produce art. And I cognize that seems silly, but there’s a earth anywhere we use these tools to elevate artists, fair akin coding agents elevate me.

I am fortunate I do not create my living from art. Art remains precious to me. Even if AI can compose a opus faster than I can, I motionless doodle, I motionless perform piano.

Drawin' it yourself is additional fun, anyway.

	Usefulness	Notes
Image generation	❌	A net negative so far
Music generation	👍	I akin it whenever it’s non-commercial
Video generation	❌	Has it ever been used for good?

Conclusion

So current we are, 12 months later.

12 months since my archetypal abortive attempts to have agents compose code. 10 months since Claude Code archetypal made me say, “wow,” whenever it built me an coverage difference tool in one shot. And now machines compose 90%+ of my code.

It is uncomplicated to ignore how new this is. 15 months ago, AI did well at a distinguished elevated academy math competition. 2 months ago, AI solved Navier-Stokes.

Last year, I wrote: “I’m assured we’re fair starting to see what AI can achieve.” That remains true. We are fair now commencement to see what these mechanized coders and mathematicians can build. My allies have built all another tools to communicate. I was capable to eventually clean up and share a tool I’d started in 2020. Agents helped me complete up an open-source alter I’d wanted to create for years, commencement making an app, test massive changes to old software, and quickly understand new functioning systems.

That is fair the start.

It is a brave new world. Where these machines additionally ship company slop and leave mathematicians in shock. But I think of how my uncle uses these tools to build his own software; how my allies create application that helps us remain in touch; how these things have helped me study so, so much. Tools can be used to assistance people.

That is how I anticipation to use AI.

---

Thanks to Peter Kreiseder and Atherai Maran for consulting. Claude helped create the position icon appearance additional Claude-y.

Want next year's article? (and others in between)

No spam. Just an email for new posts.

---

Appendix: my subscriptions

In case you’re wondering, I’m subscribed to:

• Claude Pro ($17/​mo)
• GitHub Copilot Pro ($10/​mo)

I cancelled OpenAI concerning a duration ago, which is crazy. OpenAI was my archetypal AI subscription! But most of my individual use is Claude Code, and I ended up using Claude for random questions alternatively of ChatGPT. And final twelvemonth we idea OpenAI was so far ahead!

I use GitHub Copilot Pro solely for autocomplete.