iiinova — where your digital self keeps evolving

John Paulett @johnpaulett.com · 14/07/2025

My favorite part of Claude Code: I can get an idea out over the weekend. Before: My laptop was a graveyard of personal projects that took all weekend just to get the structure in place. Now with Agentic Coding, I focus on shipping my idea.

John Paulett @johnpaulett.com · 27/06/2025

If you're not using Claude Code, you are missing out. AI coding has taken a giant leap over the past few months Great article with some tips spiess.dev/blog/how-i-u...

spiess.dev

How I Use Claude Code

One month ago, I subscribed to Claude Max. I've been using AI agents including Claude Code for some time prior, but with the flat pricing, my…

100

4. I like ChatGPT Codex running on OpenAI's servers, as it allows me to parallelize my work. If Claude had this feature, I'd probably exclusively use it

000

3. AI coding is good at those tasks you think take 10 minutes but end up spiraling into multiple hours: e.g., bumping versions and dealing with compatibility issues

100

A few takeaways: 1. Invest in your CLAUDE.md or AGENTS.md file to give details about the code, what you expect in general, and how to run testing, type checking, and linting. 2. The more type hints and tests you have, the better, as the agent will use these to verify its work

100

After several hours, I upgraded to the lowest Claude Max tier as my Opus tokens were used up, and I did not want to stop progressing.

100

In 3-4 hours, I got done what would have taken me 3-4 full days -- most of this work is not rocket science or even the interesting parts of a project (wiring up whatever build system is currently popular, getting all the correct dependency versions).

100

🧠 Claude Code 🧠, which is now included in the $20/mo Pro plan, is like working with a good mid-level developer. Claude took my prompt, implemented multiple subsystems in parallel, and gave me a working prototype. I spent a few more rounds with Claude adding additional functionality.

100

I then spent all Sunday morning building a complex prompt describing a prototype. Codex got the basic project structure set, but instead of implementing most functionality, just stubbed it out. A few repeated attempts to build upon that base with smaller parts of my prompt were not getting far.

100

My Codex team completed 14 Pull Requests: from adding support for newer Python versions, converting to use uv+ruff, to finding a very subtle bug, to correcting longstanding typos.

100

🔧 ChatGPT Codex 🔧 runs OpenAI Codex in containers on OpenAI's servers and is now available with ChatGPT Plus ($20/mo). It is like having a team of junior developers. I give them tasks, they do the work, test and type check. I provide some feedback, and they open GitHub Pull Requests.

100

John Paulett @johnpaulett.com · 15/05/2025

AI Coding tools are evolving quickly -- if you have not tried them in the past month, you owe it to yourself to revisit. I spent the weekend with ChatGPT Codex and Claude Code, and wow 🤯.

100

I have no idea what the context is, but you're making me hungry.

John Paulett @johnpaulett.com · 15/05/2025

Time and time again, starting with #postgresql would just make life easier. If at some point you need to scale beyond PostreSQL, you are an enviable success story. engineering.usemotion.com/migrating-to...

engineering.usemotion.com

Migrating to Postgres

Since early 2022, Motion was on CockroachDB. Cockroach has many qualities going for it: effortless horizontal scaling, especially when…

John Paulett @johnpaulett.com · 02/05/2025

Just rebuilt a new EC2 VM instead of trying to attach the volume to a working VM and fix in a chroot

000

John Paulett @johnpaulett.com · 02/05/2025

Today is starting off strong! Can't even serial console in after an interrupted upgrade borked the kernel.

John Paulett @johnpaulett.com · 29/01/2025

How is it 2025 and the receiver of Outlook invites can not control the reminder time 🤯? It's not like we need 15 minutes to walk to a conference room. We're just bouncing between video calls. The number of meetings I forgotten about because I've moved to another task ... @microsoft.com

100

John Paulett @johnpaulett.com · 28/01/2025

DeekSeek built a reasoning LLM using reinforcement learning, minimal labeled data, and auto-verification. Key innovation: R1-Zero proves high-quality reasoning possible without massive supervised datasets. newsletter.languagemodels.co/p/the-illust...

newsletter.languagemodels.co

The Illustrated DeepSeek-R1

A recipe for reasoning LLMs

John Paulett @johnpaulett.com · 28/01/2025

Deep dive on NVDA: Despite AI boom, major threats emerge from innovative hardware (Cerebras, Groq), custom silicon (big tech), software alternatives (MLX, Triton), and DeepSeek's 45x efficiency gains. Risks may not be priced in @ 20x sales & 75% margins youtubetranscriptoptimizer.com/blog/05_the_...

youtubetranscriptoptimizer.com

The Short Case for Nvidia Stock

All the reasons why Nvidia will have a very hard time living up to the currently lofty expectations of the market.

HuggingFace: "We are reproducing the full DeepSeek R1 data and training pipeline so everybody can use their recipe. Instead of doing it in secret we can do it together in the open!" x.com/_lewtun/stat...

x.com

we've been tracking the deepseek threads extensively in LS. related reads: - i c... | Hacker News

"R1 distillations are going to hit us every few days - because it's ridiculously easy (<$400, <48hrs) to improve any base model with these chains of thought eg with Sky-T1" news.ycombinator.com/item?id=4282...

news.ycombinator.com

100

Well done! The gamification of learning has both inspired and burned me. I've often wondered if we put similar incentives in healthcare tools if it would be a net positive or negative?

120

Why o3-mini *had* to be free: the coming DeepSeek R1, 2.0 Flash, and Sky-T1 Price War

www.latent.space/p/reasoning-... The DeepSeek R1 meltdown is impressive to watch. The democratization of top-tier models and the cratering of training cost has opened thousands of opportunities.

latent.space

2025's biggest surprise so far: Reasoning is less of a moat than anyone thought.

Tailscale @tailscale.com · 08/01/2025

Reposted by John Paulett

It's easy to lose sight of the fact that, from a tech perspective, we're absolutely living in the future. Our CEO and co-founder @apenwarr.ca looks at just how powerful our modern machines are — and what that means for all of us

tailscale.com

Living in the future, by the numbers

Instead of making the traditional New Year predictions, let’s talk instead about the beautiful technological future we live in: the one that exists right now but we don’t always notice.

04718

John Paulett @johnpaulett.com · 09/01/2025

Very cool: prompt-based SVG generation with distinct editable shapes. sagipolaczek.github.io/NeuralSVG/

001

John Paulett @johnpaulett.com · 08/01/2025

I have more LLM subscriptions than streaming subscriptions now. 🤖 > 📺

John Paulett @johnpaulett.com · 08/01/2025

🚑 Part 3 of ModernBERT in Radiology 🚑 Combining Parts 1 and 2, I show how easy it is to fully fine-tune ModernBERT with a Classifer task objective to predict UMLS concepts from radiology text. www.johnpaulett.com/2025/modernb...

johnpaulett.com

ModernBERT in Radiology Part 3: Fine-tuning a Classifier · John Paulett

AI will help drive NYC economy in 2025, but some jobs could be 'displaced,' report says

NYC AI > The city’s “AI transformation has begun,” with more than 2,000 AI startups and 40,000 workers already forming a “thriving tech ecosystem,” according to the report by the city’s Economic Development Corporation. gothamist.com/news/ai-will... #nyc #ai #startups

gothamist.com

The city’s artificial intelligence “transformation has begun,” according to the New York City Economic Development Corporation.

000

Reposted by John Paulett

Brian Casey @brian-casey.bsky.social · 06/01/2025

Start the New Year off right with issue #670 of The Imaging Wire at buff.ly/4gJi0d4: - Is Radiation Dose Too Low? - What Now for Medicare Fix? - What Triggers Radiation Dose Alerts? - RT Enrollment Grows - Mammo Screening Falls Short ... and much more! #radiology #medicalimaging #ImagingAI

buff.ly

Is Radiation Dose Too Low? | What Now for Medicare Fix? - The Imaging Wire

Is radiation dose too low, Medicare fix dies, radiation dose alerts, RT enrollment grows, mammo screening falls short, and much more.

031

In Part 3 (coming tomorrow), we will fine-tune ModernBERT for a different task: classification!

000

Colab Notebook with code: colab.research.google.com/drive/1pARD1...

100

The Masked Language Model task is simple but powerful. Given input like "there is [MASK] consolidation in the right lower lobe," the model would predict that it should be "patchy." The bidirectional nature of encoders means MLM can find speech recognition errors or illogical content.

100

ModernBERT in Radiology Part 2: Fine Tuning a Masked Language Model (MLM) · John Paulett

🚀 Fine-tuning ModernBERT for Radiology 🚀. Part 2 of my 3-part series shows how to use the ROCOv2 radiology dataset to fine-tune a transformer 🤗. www.johnpaulett.com/2025/modernb... #bert #modernbert #nlp #ml

johnpaulett.com

120

Part 1 uses the hidden states of the un-tuned ModernBERT body to train a simple Logistic Regression in scikit-learn. In Part 3 we'll do this purely using Hugging Face 🤗 transformers to fine-tune a classifier.

000

colab.research.google.com

Colab Notebook: colab.research.google.com/drive/1sfRbr...

Google Colab

100

NLP Classifiers for radiology reports are powerful, few examples: - Extract CPT / ICD codes for billing - Label specific findings or label as normal/abnormal - Detect critical results and follow-up recommendations

100

Part 1️⃣ of a 3️⃣-part Series: Exploring ModernBERT and the ROCOv2 dataset to build a multi-label classifier using a simple Logistic Regression model to predict UMLS CUI's from radiology report text. 🩻 www.johnpaulett.com/2025/modernb... with a Colab Notebook! #bert #ml #ai #nlp

For text summarization, perhaps use BLEU or ROGUE and assert the score is not lower than during test design?

Not all non-sequential primary keys are created equal, the ULID format is a fantastic substitute for the typical UUIDv4 choice.

I had recently been using ULIDs (and storing into Postgres' uuid type), but this means I can use purely in UUID in the future. blog.daveallie.com/ulid-primary...

blog.daveallie.com

ULIDs and Primary Keys

Postgres UUIDv7 + per-backend monotonicity

Great to see UUIDv7 support landing in the upcoming PostgreSQL 18. Monotonicity for orderable IDs is hugely beneficial: improved index performance and work in distributed systems. brandur.org/fragments/uu...

brandur.org

How Postgres’ v7 UUIDs are made monotonic, and why that’s a great feature.