I’ve experienced this at work because I’ve observed agents being able to do more and more sophisticated work and work for longer periods of time.

I’ve experienced this personally by attempting to create different web tools, videos or even 3-D animations recently. Each time I’m able to literally just create what I had wanted and it’s often better than I expected.

I created a web app to track different definitions of AGI and see if we’ve met any definitions yet. First, the fact that Claude Code + Opus 5.5 could do the research, create this application for me and then deploy it is unreal. I would not have attempted this type of project pre-2023. I’d have these questions but those questions would just lie around in my Todoist. Going back to definitions of AGI, only one definition, Peter Norvig et al.’s, has been met:

“The standard is breadth, not reliability or expertise. The frontier models of 2023 already qualified, and today’s models qualify more clearly. It is the only definition fully met, because it asks the least.”

That, in my estimation, is a little early. The frontier model of 2023 was GPT-4 and there were no coding harnesses like Claude Code or Codex yet. Yes, the frontier models of 2023 could answer many questions; however, they were limited in how much actual work they could do in actual systems. Now in September 2026, with the increase in model intelligence, improvements in model harnesses and the types of tools available, the range of tasks agents can do has exploded. I think that’s why you see an explosion of assistants ever since OpenClaw in December 2025. I’ve been using Claude Code + GitHub as my personal assistant. Now there’s Grok Bot, Muse and today I heard about Dots from OpenAI and Fo from Wajo.ai. I wonder if Anthropic will release a similar personal assistant product too.

On that note, I was curious how AI has progressed the last 3 years on axes like model intelligence and agent harnesses. I was imagining two timelines with different points marking key developments. After some iteration, I wound up building a separate application where you can see the timeline of development from 2023 to Sep 2026 across:

• models
• harnesses
• achievements
• research
• compute

Claude also made a video walking through the AI progress the last few years that’s pretty good.

But both the AI progress web app and the creation of the video are yet another example of an AGI-moment for me. Creating these types of apps was not possible before. It was possible given enough time. But now, given enough tokens, one can create almost anything. Creating has become a new way of learning and exploring my interests. It’s also a new form of entertainment for me. I find it much more rewarding and exciting to learn, create and explore something than, say, watch a movie nowadays.

My favorite Bible is the ESV Study Bible. Two of my favorite “features” of this Bible are the maps and the illustrations. The temple illustrations are still so amazing to me like this one of Solomon’s Temple. I tried using Blender to build 3-D animations of various temples in the Bible like Solomon’s or Herod’s. I’ve never used Blender before. But Claude or any frontier model with an agent harness is able to write code to drive arbitrary software to do anything. Creating 3-D visualizations is no different. A few more prompts later I have 10+ places I can explore along with walk-through videos. Here’s Solomon’s temple:

And its walk-through video:

Even in writing this post, Claude is helping me add the temple screenshots that I used to add manually. This is the command it used to take a screenshot using a headless Chrome.

"/Applications/Google Chrome.app/Contents/MacOS/Google Chrome" \
  --headless=new \
  --use-angle=swiftshader \
  --enable-unsafe-swiftshader \
  --ignore-gpu-blocklist \
  --hide-scrollbars \
  --force-device-scale-factor=1.5 \
  --window-size=1440,900 \
  --virtual-time-budget=20000 \
  --screenshot=solomons-temple.png \
  "https://apps.lawrencewu.net/sacred-spaces/#site=temple&stop=0"

I actually didn’t know you could do that. Though it is similar to what Simon Willison has in shot-scraper. Sometimes I still like to read the model’s thinking tokens and read some of the commands it’s running to learn.

It’ll be fascinating to see how AI continues to develop. But we’re at the point where the potential applications and use-cases for AI are exploding. I’m interested to see what kind of completely new use-cases, businesses or ideas are unlocked with cheaper and faster intelligence or even other models like the one that comes after Jev.

Some things that absolutely amazed me:

• When he ejected, his parachute didn’t deploy. He was free-falling and he hit the ground at an estimated 70-100mph yet he survived!
• He had a broken back, broken arm, broken shoulder, sprained ankle and cuts and scrapes to his head and face
• He knew the Iranians were going to be looking for him. Bravo landed in a mountainous region of Iran. He thought higher ground would be safer so, despite his broken bones and other injuries, he started climbing this 7,000-foot mountain. He said “never let a lack of motivation put you on Iranian TV.”
• He had no basic survival supplies. Normally they’re in a package under the seat but because the seat was damaged all he had was minimal survival gear in his vest.
• Iranian search parties were actively hunting for him and probably came within a few hundred meters of him.
• The first message he sent back to the US command was:

“I’m hurt, I’m moving, God is good.”

The above message was so moving. Despite the situation he was in and every reason to complain, he knew confidently God was still good and sovereign.

I’m reminded of the importance of:

• faith: during times of dire need
• training: also during times of dire need
• how both faith and training are important
• not panicking, not worrying about things that are outside of your control, prioritizing the problems ahead of you and then simply solving the most important problem that faces you

This is one of my uses of Workflowy: a taxes section under finances. I’d create a new list for each year. Any time something came up that was tax related, a deduction I wanted to remember for example or an I-Bond sale when I used to own them, I would make a note. Then come tax filing time, taxes were quite pain free, possibly even enjoyable?

I tried switching to Obsidian last year because I was intrigued at storing my notes in plain text and markdown. I knew agents would work more effectively with markdown. But it didn’t click for me. One of the reasons was I tried also using AI to create notes and manage my Obsidian for me. That got me all confused because part of the value of a human-managed PKM is you know where things are AND you know how things are worded. It’s like your home. If you put something away, you’ll be more likely to remember where it is. When the AI started filing things for me, I was somewhat at a loss at finding information.

I recently read the book How to Take Smart Notes by Sonke Ahrens. He writes about the idea of the “slipbox” which is the German translation of zettelkasten. The general idea you take notes about things you read/listen to/watch. This is the capture step. I did this before in Workflowy so this isn’t new to me. But the next step is key. At another time, usually at the end or beginning of a day, you review these notes, what you’ve learned and write a new note, promoting an idea into something the “slipbox” where you codify the learning. You also have to make connections with other ideas in the slipbox. On a side note, this is one of the things I love about Obsidian, how easy it is to make links between notes. I could do this in Workflowy but it wasn’t a native feature at the time. In Obsidian, I can type two “[” and I can quickly find notes to link to. Though I just checked and Workflowy now has this feature too.

For example in my slipbox, I have a note called Memory is important for both humans and AI. I connected it to 2 other slipbox notes. One was about AI capability compounding because of models, harnesses and skills all improving together but external memory is another potential axis of compounding improvement. This note being my slipbox made it so I came across it a few times after I had written it. And it led me to consider writing this series about memory.

It’s been a few weeks since I’ve started writing notes into a slipbox. I can’t say I’m used to it yet but I do see the value of it. The author of How to Take Smart Notes was inspired by Niklas Luhmann. Luhmann was a German sociologist who used a slipbox. Amazingly it was a physical slipbox with actual index cards. Over his career he built a zettelkasten/slipbox of 90,000 cards! People say his slipbox is what allowed him to publish 70 books and 400 scholarly articles across a vast array of subjects.

This is a picture of his slipbox. I still can’t believe he hand maintained the links between the physical cards!

I imagine using Workflowy over a 10 year period has shaped my mind a certain way and allowed me to remember certain things. Simply writing down things, reflecting over things, learning where things were so I could retrieve them probably did change me. Now, as I’ve used Obsidian more and specifically writing notes into the slipbox I’ve observed being able to make some connections across things I’ve been thinking about like: AI, theology, personal finances and homeschooling.

As I reflect on the different amounts of complexity of intaking information and organizing it, there are levels:

1. Little or no reading
2. Reading but not writing it down, no PKM
3. Simple PKM, somewhat scattered though
4. Structured PKM by categories or areas of responsibility
5. Connected PKM like a slipbox, writing slipbox notes

I had listed writing in my top 10 human skills. Growing up I didn’t enjoy writing. AP English and AP Literature were a slog. A few years ago, I started Phase 2 of the ACBC Biblical Counseling certification which requires writing 40+ essays. Through writing those essays, I grew to appreciate the practice of writing. It required researching a given theology or counseling topic, finding Scripture, listening to lectures, writing notes down, organizing the notes into an outline and then turning the outline into a 1-1.5 page essay. AI almost certainly can write a better essay than me but that wasn’t the point. The goal of writing the essays was to be changed: to understand these diverse set of topics and formulate an opinion about it. Using AI would’ve shortchanged me completely.

Using AI to write and organize my notes in Obsidian wasn’t a good idea for that same reason. I was depriving myself of one of the best tools for thinking, writing. I’m excited to keep writing notes and continuing this new practice of writing slipbox notes and seeing what comes out of it!

Fast forward to June of this year, I was thinking of creating an agent to help us manage our homeschool. There are so many moving pieces in a homeschool:

• curriculum
• plans / pacing
• state standards

Things are pretty dynamic too depending on how the student progresses, often plans need to be updated. I thought AI could somehow help this.

Two weeks ago, I saw YC released qm, yet another agent harness. If 2024 was the year of agent frameworks, then 2025-2026 has been the year of agent frameworks with so many coming out like Pi, OpenCode, Goose, Omnigent, etc. I spent some time research with Fable some options:

I even considered setting up Slack and using Claude Tag given the usefulness of it I’ve seen at work.

In the end I settled on simply:

• Agent Harness: Claude Code
• Knowledge Base: GitHub

It’s been only a couple weeks but it’s been surprisingly useful. Some of the benefits so far:

• My wife and I being on the same page - planning and plans are now documented well and can be read, not just in one person’s brain
• Creating schedules/checklists
• As flexible as we want it to be
• Cataloging / organizing our books at home
• Keep track of: work for each day, books read, etc.

I never imagined needing to teach my wife about GitHub, pull requests and the difference between a model and an agent harness :)

I like Claude Code as the interface because I can interact with my homeschool repo on mobile or desktop. Sessions from mobile are also viewable on the desktop. I don’t have to worry about infrastructure or services dying. I chose GitHub as the knowledge base because I figured Claude Code has been trained to interact with GitHub repos and knows how they operate, there are existing conventions like README.md and pull requests / diffs that can be repurposed for this non-coding use case.

For example this is an example of a query, “Can you give me an overview of this upcoming homeschool week?” And what happens:

• Set up a cloud container
• Cloned repository
• Started Claude Code
• And it proceeds to do the typical (but still surprising and amazing) agentic search by exploring the repo in order to answer the question

I imagine a similar pattern can be done for other use cases like:

• a financial agent
• a trip planning agent
• a generalized personal assistant
• an assistant at work

If you’re interested, this is a Claude Artifact I generated that describes how the homeschool agent operating out of the repo works. You could just point Claude to it to set it up. This is a starter repo that contains a skeleton of files that we use for our homeschool agent. I taught some of my friends over the weekend how to set this up for their homeschools, it’ll be interesting to see whether this type of setup (Claude Code + GitHub) will be manageable for people who are using Claude Code and GitHub for the first time.

Note this solution costs $20/month (Claude Code Pro plan). I was curious about Grokbot from Cursor/SpaceX but had no idea it requires the $259.99/month Ultra plan!

Given the Mac I already own, what is the best model I can actually run?

That question gets surprisingly complicated. A model name alone is not enough. You also need to know which quantization you are downloading, how much unified memory it needs, whether the benchmark score belongs to that exact artifact, and whether your inference engine and coding harness can work with it.

So I built Local AI, an interactive map of model intelligence versus local model size.

I wanted to know how crazy, so I started ranking the offseason’s signings by a metric I’m calling surplus cap points — the gap between a player’s estimated fair salary and what he actually signed for, as a share of the salary cap. Then I kept going and built a second leaderboard for the best bargains of the salary-cap era, which is where Curry’s $44 million extension and Chauncey Billups’ mid-level deal live.

At that point it wasn’t a blog post anymore, it was a data app. So that’s what it is now:

The Bargain Bin →

Both leaderboards are there — the 2026 offseason and the all-time steals — with the methodology, the sources, and all the caveats. The short version: LeBron at $4 million a year is the value signing of the offseason at 15.8 surplus cap points, more than double the next-best deal. And if he plays to even the conservative projection behind that number, the contract lands in the same neighborhood as the bargains that underpinned the Bulls and Pistons championship cores.

One housekeeping note: I’m keeping this blog for writing. The interactive stuff — this one, the Bible apps, whatever comes next — lives at apps.lawrencewu.net, and posts here will just point to it.

“Sorry, there’s no slides. I have no slides.”

No deck, the whole time. Just Garry, a mic, and 20 minutes of dense material about AI-native orgs, skill files, and “company brains.” I thought it’d be a fun experiment to see if Claude Code could reverse-engineer a slide deck purely from the talk itself — no slides to work from, just the words.

Claude Fable 5 was set to start drawing from usage credits on July 13, but we’ve extended our promotion period and it now stays in your plan through Sunday, July 19 at 11:59 pm PT.

I’ve used around $1,000 of Fable 5 tokens in the last 7 days on my subscription plan.

I’ve had fun pushing Fable 5 to its limits. I think what it’s pushed me to do is to be more ambitious with my ideas and projects.

For example I had a backlog of ideas I’ve been compiling over the last 5+ years. I haven’t had time to work on any of these things. Thinking Fable could help out, I just exported the list as a CSV from Todoist and had Fable rank order the projects and start implementing them. I was pleasantly surprised that Fable did a pretty good job on most of them.

It’ll be interesting if Anthropic pushes back the date again this weekend. With GPT 5.6 Sol being competitive with Fable, I expect if OpenAI keeps Sol in the OpenAI subscription, Anthropic will likely hae to do the same with Fable in order to stay competitive with consumers.

The raw view count alone is a little unfair because the uploads are staggered. A talk posted on June 9 has had a month to accumulate views; a talk posted on July 8 has had about a day. So the main ranking below is:

views per day = views / max(1, days since upload)

As of August 1, 2026 at 9:54 PM Pacific, the early winners are:

1. Why Agentic Systems Need Ontologies — Frank Coyle, UC Berkeley
2. Field Guide to Fable — Thariq Shihipar, Anthropic
3. Harness Engineering is not Enough: Why Software Factories Fail — Dex Horthy, HumanLayer
4. Understanding is the new bottleneck — Geoffrey Litt, Notion
5. “The engineer of the future is the person who is able to choose what is worth doing.” — Addy Osmani

uv tool install yt-dlp

import datetime as dt
import json
import subprocess

CHANNEL_URL = "https://www.youtube.com/@aiDotEngineer/videos"
REFERENCE_DATE = dt.date(2026, 8, 1)
PLAYLIST_END = 250

EXCLUDE_TITLE_SUBSTRINGS = [
    "Vibe Reel",
    "Things to Know about AIE",
]

cmd = [
    "yt-dlp",
    "--playlist-end",
    str(PLAYLIST_END),
    "--skip-download",
    "--ignore-errors",
    "--print",
    "%()j",
    CHANNEL_URL,
]

proc = subprocess.run(
    cmd,
    check=False,
    text=True,
    capture_output=True,
    timeout=420,
)

rows = []
for line in proc.stdout.splitlines():
    if not line.startswith("{"):
        continue

    video = json.loads(line)
    title = video.get("title") or ""
    upload_date = video.get("upload_date")
    view_count = video.get("view_count")

    if not upload_date or view_count is None:
        continue
    if upload_date < "20260608":
        continue
    if any(skip in title for skip in EXCLUDE_TITLE_SUBSTRINGS):
        continue

    uploaded = dt.datetime.strptime(upload_date, "%Y%m%d").date()
    days_since_upload = max(1, (REFERENCE_DATE - uploaded).days)
    views_per_day = view_count / days_since_upload

    rows.append(
        {
            "title": title,
            "url": video.get("webpage_url")
            or f"https://www.youtube.com/watch?v={video.get('id')}",
            "upload_date": uploaded.isoformat(),
            "views": view_count,
            "days_since_upload": days_since_upload,
            "views_per_day": views_per_day,
            "duration_minutes": round((video.get("duration") or 0) / 60, 1),
        }
    )

rows.sort(key=lambda row: row["views_per_day"], reverse=True)

print("| Rank | Video | Uploaded | Views | Days | Views/day |")
print("|---:|---|---:|---:|---:|---:|")

for rank, row in enumerate(rows, start=1):
    print(
        "| {rank} | [{title}]({url}) | {upload_date} | "
        "{views:,} | {days_since_upload} | {views_per_day:,.1f} |".format(
            rank=rank,
            **row,
        )
    )

I gave the same open-ended prompt to both models and let each build an interactive 3D knowledge graph of the Bible — 66 books, tens of thousands of Treasury of Scripture Knowledge cross-references, and curated theology/people overlays. I had come across Marble’s curriculum map and was blown away by it. This is the final deployed app

The two models I used:

• Sol 5.6 High — OpenAI’s gpt-5.6-sol at high reasoning, driven by the Codex CLI.
• Fable High — Anthropic’s claude-fable-5 at high effort, driven by Claude Code.

Same starting prompt, same source data, two independent one-shot sessions.

The prompt was the same: > I’m interested in building a visualization that is like https://withmarble.com/curriculum/. The data is here: withmarbleapp/os-taxonomy. I’d like to build it for a Christian use case: Bible will be the star, verses, books, cross references and secdonarily topics in systematic theology or people in the Bible. I put the tsk data in the folder too you can use.

This my Claude desktop app:

This is my Codex desktop app:

Here’s what each cost and how each worked — and why I ended up deploying the more expensive one.

Sol   (61 tool calls):  ████████████████████████████ 61 shell
Fable (153 tool calls): ████████████ 39 Bash
                        ██████████████████████ 66 browser (46 eval + 11 shot + …)
                        ████████████ 36 edit/write   █ 5 read

The comparisons to Opus 4.7 and GPT 5.5 were quite strong.

I was surprised Anthropic didn’t put Mythos Preview as a model they are comparing against. Naturally I was curious whether Opus 4.8 was better than Mythos Preview (which was released almost 2 months ago on 2026-04-07).

I went back and looked at the Mythos Preview System Card and saw these two tables on page 187-188 that show Mythos Preview’s performance on these common benchmarks.

I then asked Claude Opus 4.8 (in Claude Cowork) to add Mythos Preview and Opus 4.6 scores for SWE-bench pro, Terminal-Bench 2.0 (though 4.8 is using Terminal-Bench 2.1) and Humanity’s Last Exam to the same table.

It generated a SVG and then converted it to a PNG. The code for the SVG is here. This is the result:

I’m pretty happy with the result. I don’t have Anthropic’s font handy to make it exactly the same but figured it’s close enough. Opus 4.8 still lags Mythos Preview. It’s pretty amazing there is still room to improve these large language models. There are rumors that a Mythos class model will be available soon.

I love reading technical blogs at other companies. I’ve learned so much from blogs especially ones from Anthropic and LangChain. It’s exciting that we’re starting one at UKG. One of the posts, Foundation Models for Workforce Management showcases some of the work my AI team at UKG has been working on.

I believe most companies that have large datasets can train these types of domain-specific models that use domain specific vocabulary/tokens. Stripe for example has done something similar in the payments space. The article talks about a 2-stage training phase:

1. Phase 1 - Pretraining: We first train the Workforce Foundation Model to predict what tends to happen next in a worker’s behavioral sequence. In plain English, the model reads large amounts of historical labor data and learns common workforce patterns.
2. Phase 2 - Task Adaptation: We then freeze the pretrained model and train LoRA adapters for a downstream problem such as overtime forecasting.

We trained a single model across 5 retail customers on 12.6 million rows of data. This workforce foundation model outperformed a traditional model on 4 out of 5 of the customers and reduced average prediction error by 2.7%. What was even more amazing though was the model was predictive on tenants that were not part of the training data! On one unseen tenant the workforce foundation model even outperformed a traditional ML model trained on that customer’s data.

I don’t think my team and I would have been able to undertake this research without coding agents like Claude Code. We could go way deeper in an area we were familiar with: machine learning and deep learning.

The talk covers the Agentic AI Stack — the mental model I use to think about how humans and AI systems work together — and practical strategies for becoming a more effective knowledge worker in the age of agents.

The diagrams and materials from this talk are here.

• Agent Skills Docs
• Matrix Scene
• Anthropic Knowledge Work Skills

No coding background required. If you’ve used ChatGPT, you’ll understand this.

The slides and materials used in the video are here.

In Chinese, Good Friday is 耶穌受難日 (Yēsū shòunàn rì), literally “The Day of Jesus’ Suffering” and Easter is 復活節 (Fùhuó jié), literally “Resurrection Holiday”. I appreciate how the names of the holidays in Mandarin capture the essence of the two days: Jesus’ suffering and Resurrection.

Jesus suffered through his death on the cross and He resurrected. Two realities that Christians around the world commemorated this weekend. Both are core doctrines of Christianity. But why would we care so much about a person dying on a cross? And why should we possibly believe someone resurrected? tl;dr - it is through faith in Jesus’ death and resurrection that our sins are forgiven and eternal life / salvation is given.

I didn’t consider these things until college. I’m still thankful to God for those seniors who lived in my dorm who shared the gospel with me. I learned the cross was significant because it was through Jesus’ death on the cross that man’s sins were atoned for. Sin is breaking God’s law. Everyone has in some shape or form broken God’s law. Deep down we know it. As a non-Christian, I knew it. I became more painfully aware of my sinfulness as I read the Bible and learned about sins of greed, selfishness, lust and anger. As I became more aware of the sin problem, I desperately sought a solution. That’s when the gospel finally became “good news”, Good Friday became good and I saw Christ’s sacrifice as not only significant but personally relevant for me.

Galatians 3:13 summarizes this well:

Christ redeemed us from the curse of the law by becoming a curse for us

The resurrection is very significant because without it, Jesus would just be someone who made these extraordinary claims (that he could forgive sins and that He was God) and his body would still be in the grave. The resurrection is so significant that if it isn’t true, the Bible actually says preaching and our faith is futile and a waste. Any person could have made the same claims Jesus made (though not everyone could have performed the miracles Jesus did). Jesus’ resurrection validates what he taught, proves he was in fact God and that he had power over death.

As a non-Christian, I remember asking my friends who were Christian and other Christians I met, “Wait, so you actually believe someone came back from the dead? How is that possible? What evidence is there?” During that time, someone shared this argument from Gary Habermas called the Minimal Facts Argument that helped me quite a bit. He identifies 6 historical facts where each of them is confirmed by strong and independent sources, some even critical scholars. The facts were:

1. Jesus died by crucifixion
2. His followers soon afterward had real experiences that they thought were actual appearances of the risen Jesus
3. The disciples lives were transformed as a result, even to the point of being willing to die specifically for their faith in the resurrection message
4. These things were taught very early, soon after the crucifixion
5. James, Jesus’ unbelieving brother, became a Christian due to his own experience that he thought was the resurrected Christ
6. The Christian persecutor Paul (formerly Saul of Tarsus) also became a believer after a similar experience

Once I believed these 6 things to be historically true, it came down to identifying what reality best explains these facts. Some people say Jesus fainted (swoon theory) but this contradicts the first fact. Some people say Jesus had a twin (twin theory) but there wasn’t much evidence for this. Some people say the disciples stole the body but it’s doubtful the disciples would be willing to be killed for a lie. The resurrection was the only thing that explained all of the minimal facts.

If you’ve never considered these things, I hope you will at least spend some time trying to answer those questions of why did Jesus suffer? And did Jesus really resurrect? And if these things really did happen, what are the implications on my life? I’m happy to have a conversation if you have any questions too.

The essence of sin is man substituting himself for God, while the essence of salvation is God substituting himself against God and puts himself where only God deserves to be; God sacrifices himself for man and puts himself where only man deserves to be. (John Stott)

If Jesus rose from the dead, then you have to accept all that he said; if he didn’t rise from the dead, then why worry about any of what he said? The issue on which everything hangs is not whether or not you like his teaching, but whether or not he rose from the dead. (Tim Keller)

Every year I try to read through the Bible passages that correspond to Jesus’ last week he walked on the earth. Today I was thinking about Matthew 27:46:

And about the ninth hour Jesus cried out with a loud voice, saying, “Eli, Eli, lema sabachthani?” that is, “My God, my God, why have you forsaken me?”

Jesus is quoting David’s words in Psalm 22:

My God, my God, why have You forsaken me?
Far from my salvation are the words of my groaning.
O my God, I call by day, but You do not answer;
And by night, but I have no rest.
Yet You are holy,
Enthroned upon the praises of Israel.

Jesus is quoting Psalm 22:1 while hanging on the cross. David says this because he feels this intense sense of personal abandonment. Jesus feels this too but in a much greater way. Jesus not not only went through intense, unimaginable physical pain but arguably more intense, also unimaginable spiritual pain by being abandoned by the Father.

As a Christian, I will never have to go through what Christ went through because He went to the cross, experiencing the divine judgment I as a sinner deserve and should have experienced. Praise God. The Bible reassures me when it says nothing will be able to separate me from the love of God, which is in Christ Jesus our Lord (Romans 8). How did Jesus show that love? On Good Friday.

But God demonstrates his own love for us in that while we were still sinners, Christ died for us. (Romans 5:8)

I created a better version of a tutorial called a Deep Dive into Coding Agents like Claude Code.

The diagram and markdown file shown in the video are here.

He originally wrote it improve large language model training in his nanochat repo. But people figured out the general pattern can be used to optimize and improve anything. It works by

• optimizing a single metric
• in short training loops so 50-100 experiments can run in a short amount of time
• all of this is done in an agent loop to take advantage of the intelligence of a model

It is absolutely trivial to use, just prompt Claude Code with:

use karpathy's autoresearch automatically to improve this model

I say this is a killer app for coding agents because without them, autoresearch wouldn’t work. You need both the language model for it’s reasoning capability to know what to do, what research steps to take, what model parameters to tune and you need to have the agentic harness to write the code, run the code and use tools. It’s amazing to see this all come together with Claude Code and autoresearch.

There’s already an awesome-autoresearch repository collecting some of these examples.

I’ve used it for models at UKG improving models we’ve trained by 15-20% - models I’ve worked on for months. The general problem with why we didn’t see these improvements is the cognitive load and time to try different parameters. A few years back there was always the promise of AutoML, this is probably the first example of truly automated machine learning.

I’ve attempted to try to push toward automating the whole ML lifecycle in agentic-ml-plugin but autoresearch is by far the best implementation I’ve seen for the training/finetuning piece.

I’m grateful for the ability to write. I have really come to see the value of writing more and more. Writing is in my top 10 list of human skills in the agentic knowledge work era.

I thought I’d pick out a couple of my favorite posts from each year:

Here are the top 10 human skills I think are currently most valuable in the era of agentic knowledge work. These are somewhat sorted in decreasing order of importance but they’re all important.