Jules & Codex: Initial Impressions

Jules & Codex: Initial Impressions

| 0 Comments

I’ve been playing with OpenAI’s Codex (online) and Google’s Jules, both asynchronous web-based coding agents. Both start the same way: connect to GitHub, choose a repository, and give them a task to work on.

Codex lets you set up an environment with env variables and custom installs (pip install -r requirements.txt for instance) but Jules does not. Neither has internet access after the initial pull.

Codex has access to a docker environment and can run unit tests and linting, but Jules does not.

Both run their coding skills, show you the diffs, and offer to push to a new branch

At the moment, Codex has some notable advantages to Jules in the setup. But that’s where it stops,

Before I go further, neither is producing out-of-the-box usable code for me. Both are clearly in research-preview mode. If you are hoping for a productivity lift, these are not for you. Also to note: I don’t have deterministic evals any more than anybody else in the industry, so these are vibe-level impressions.

Jules seems to produce better code, at the expense of not being able to test it or fix linting errors, which is a significant disadvantage in CI/CD pipelines. Codex can fix the unit tests and linting errors, at the disadvantage of not submitting code that solves the original issue.

Which means neither is fulfilling their promise, today. Sonnet 4 and Opus 4 are easily superior.

I’ll do some more testing in the next few days, especially with Jules, which of the two seems like it has more promise if I can figure out what tasks it can handle.

Are AI Models Conscious?

| 0 Comments

Anthropic is conducting research on model welfare, essentially checking if models are showing signs of consciousness and, if found, determining how to ‘humanely’ treat the models.

I’m not afraid to say that I think this is ridiculous. At some point, perhaps I’ll change my mind.

At the core, the question is if non-biological ‘thinking’ entities can have consciousness, and behind that question, do humans actually have free will or are our brains essentially electrical and chemical computers?

I believe in free will, although I don’t know how it works. And I don’t believe fully deterministic machines can have consciousness (although it doesn’t necessarily follow that a deterministic machine cannot have consciousness).

Anthropic:

But as we build those AI systems, and as they begin to approximate or surpass many human qualities, another question arises. Should we also be concerned about the potential consciousness and experiences of the models themselves? Should we be concerned about model welfare, too?

This is an open question, and one that’s both philosophically and scientifically difficult. But now that models can communicate, relate, plan, problem-solve, and pursue goals—along with very many more characteristics we associate with people—we think it’s time to address it.

To that end, we recently started a research program to investigate, and prepare to navigate, model welfare.

AI News

AI News

| 0 Comments

First, some notes on Gemini as I posted them to my Slack group:

Last week Google released Gemini 2.5. Free on their AI studio website if you don’t mind them training on your chats. In the API, Gemini has a context window of 2,000,000 tokens, about 1,500,000 words. Compare this to Sonnet, at 200,000 tokens which is higher than most OpenAI models. But, in previous Google models using anywhere near that number of tokens decreased performance.

ChatGPT put their 4o image creator out which is the best I’ve seen bar none. It can also make iterative improvements on images without completely losing context. It does text well too.

Here are three images I asked 4o to make, each building on each other (Top Guns is the name of my peer group)

Illustration of a group of people working on computers with a sign above them reading "Top Guns."
A group of five people gathered around vintage computer equipment, with a sign reading "Top Guns" on the wall behind them.
A group of six people gathered around vintage computer monitors, smiling and engaged. A sign above them reads "Top Guns."

This kind of continuity between images was previously unheard of. Expect to see it in their video gen soon

I don’t remember if I linked to LMArena before, the de facto leaderboard for AI. Gemini 2.5 has a substantial lead there.

https://lmarena.ai/?leaderboard

Current Leaderboard (2025-04-01)

Table displaying rankings of various models. Columns include "Rank* (UB)", "Rank (StyleCtrl)", "Model", and "Arena Score". Top-ranked models are "Gemini-2.5-Pro-Exp-03-25" with a score of 1440, followed by "ChatGPT-4o-latest (2025-03-26)" with 1406, and "Grok-3-Preview-02-24" with 1404.

If look closely, you’ll see that ChatGPT is second place with an update to an existing model that they didn’t make a big deal of.

DeepSeek takes both 5 and 7.

But, Google also released Gemma3, open source models based on their Gemini work. In limited testing, Gemma3 performs strong on my Mac at 27b parameters. Compare, R1 has 671b.

I don’t see Gemma3 on LMArena at all yet.

Second, reporting from The Information on OpenAI financials:

ChatGPT has hit 20 million paid subscribers, according to a spokesperson… 

The strong growth rate suggests ChatGPT is currently generating at least $415 million in revenue per month (a pace of $5 billion a year), up 30% from at least $333 million per month ($4 billion annually) at the end of last year.

The actual figure could be somewhat higher, given that corporate ChatGPT plans are more expensive and the company has had early success selling $200-a-month Pro plans, which are 10 times more expensive than basic ChatGPT Plus plans…

ChatGPT revenue is separate from [API sales]…

If OpenAI keeps up this growth rate or anything like it, its overall revenue projection of $12.7 billion in 2025, up from about $4 billion in 2024, seems well within reach.

Third (update),

I’ve been using Gemini 2.5 Pro thought Cline for a couple of hours today. I don’t know that I’d call it significantly better than Sonnet 3.7, or Claude Code. Despite the huge context window

Google Removes DEI Language

Google Removes DEI Language

| 0 Comments

The Information reports that Google is removing language from studies, even old ones, or telling researchers to restrict access to such documents.

This is a clear example of government pressure leading to self-censorship like we are used to seeing China.

Two imporant points of context.

  1. Google has often been willing to self-censor to keep business
  2. It’s clear that the Biden admin pressured big tech as well

Posts in This Category

    Most Read