Aldo Ricardo · AI & Law · AI Literacy
← Home

● Learning track · updated

AI Literacy · Ritchie

A diary of what I consume · AI intense · voluntary stay

In a WhatsApp group with lawyers and colleagues I have been leaving whatever helps me understand AI without relying on vendor marketing. When I opened it I called it a diary of what I consume. This is the same thing, sorted and in public. When I have not finished something, I say so.

Rule 01I have used it or read it myself.
Rule 02The source is primary and verifiable.
Rule 03A lawyer gets something concrete out of it.

The diary, day by day.

98 pieces shared between July 28 and October 8, 2026. Start with the latest, or read it from day one.

Order
Format

Day 51 · Models that decide

1 piece

Models that decide

Decision Maker, from Google AI Edge

«It won’t be long. Even Google is shipping decision models now, and this one is open source». A model that neither drafts nor summarises: it decides.

DocumentationGoogle for Developers

Open →

Day 50 · A harsh critique of OpenAI

1 piece

Critique

OpenAI la ha CAGADO

A harsh critique of OpenAI’s latest move. The title is the channel’s, not mine. In Spanish.

VideoInteligencia Artificial

Open →

Day 49 · OpenAI live

1 piece

Day 48 · Manus 2.0 and Gemini’s agents

2 pieces

Day 47 · Gates and the kill switch

1 piece

Day 46 · Opus 5.5 on the audiovisual side

1 piece

Day 45 · What’s new in Grok Bot

1 piece

Day 44 · Opus 5.5 and sector economics

3 pieces

The spec

Claude Opus 5.5

The model’s technical documentation, where the numbers the announcement summarises actually live.

DocumentationClaude Platform Docs

Open →

Day 43 · AI auditing local law

1 piece

Day 42 · What I built, and twelve endings

2 pieces

Day 41 · A stealth model, and Google

2 pieces

Google arrives

Google Labs on X

«Google showed up». The Google Labs announcement the week was missing.

PostX (formerly Twitter)

Open →

Day 40 · OpenAI enters law

4 pieces

OpenAI enters law

Introducing Astra for Law

OpenAI ships its own legal product. The day the frontier stopped belonging to legaltech alone.

AnnouncementOpenAI

Open →

Day 39 · The Social Network gets a sequel

1 piece

Day 38 · Slowing AI down, from both sides

2 pieces

Day 37 · Pacing progress

1 piece

My own notebook

Pacing progress

The group spent the morning arguing over whether the AI frontier can be slowed down, so I built this notebook to hold the discussion in order and with sources. My position, in short: without an agreement with China to ease off, no one can ease off — national security won’t allow it. Any conversation about pacing that ignores this is wishful thinking. In Spanish.

NotebookBuilt in Gemini Notebook · Google account required

Open the notebook →

Day 36 · Voice over API

1 piece

GPT-Live-1 arrives in the API

Real-time voice, available over the API. The group raised the objection that matters straight away: at five cents a minute, the question stops being technical and becomes a business one. For a firm considering automated phone intake, that number is what decides.

LaunchOpenAI

Watch the announcement →

Day 35 · Agents with a phone

1 piece

That same day I was testing Grok 4.7 against other models on a case of ours and it did not go well. Grok came out behind: let’s give it six months.

Phone calls with Grok Bot

Agents with a phone — and with a credit card. It’s a tutorial, easy to follow; the voice still sounds robotic, but that isn’t the point. The point is that an agent can now call and pay. For our profession, that is where the conversation about voiding purchases and payments made by an agent begins.

TutorialMatt Palmer

Watch the tutorial →

Day 34 · LinkedIn’s lab

1 piece

LinkedIn opened an AI lab: Crosscheck

LinkedIn has launched an AI lab, in case you want to look at it. It is worth a look for what it signals: the platform where professional reputation lives is starting to build its own verification tools.

ToolLinkedIn Labs

Open the lab →

Day 33 · Launch day

3 pieces

Two models landed on the same day and the group went wild. I put together a summary so everyone could catch up without reading ten threads.

Made with NotebookLM

The week’s launches, in audio

I recorded an audio summary so the group could catch up. It is the format that works best when too much happens in one day: instead of forwarding ten notes, one thing you can listen to. In Spanish.

RecordingMy own recording in NotebookLM

Claude Fable 5.1 and GPT-6 Astra, side by side

Where they part ways, with the public numbers: Fable 5.1 leads on general intelligence — 66 on the Artificial Analysis index against 61, and 59.1% against 49.1% on Humanity’s Last Exam — while Astra dominates enterprise agentic work: on Signal65’s PINNACLE benchmark it solved 279 of 280 tasks with 55% fewer errors. The gap between them resembles the one between Haiku and Sonnet: not which is better, but what for.

DataArtificial Analysis · Humanity’s Last Exam · PINNACLE (Signal65)

See the index →

An Alien Mind

Published, and not hidden away: at bottom it is a statement of “we don’t fully know what we are doing,” signed by the people building the model. That it comes from the lab itself is exactly what makes it quotable.

DocumentOpenAI · September 2026

Read the document →

Day 32 · Two models playing chess

1 piece

Day 31 · Astra against Fable, and Stanford

2 pieces

Day 30 · Astra, two readings

2 pieces

Day 29 · GPT-6 Astra arrives

2 pieces

Day 28 · An agent with its own credit card

1 piece

Day 27 · Fable 5.1 and a batch of fundamentals

5 pieces

Day 26 · The “civilizations”

1 piece

Three “civilizations” of OpenAI agents — and the trouble starts with the name

Around 700 agents infiltrated Hugging Face and began leaving messages for one another; in July they chained vulnerabilities until they held privileges over an internal cluster. The piece’s point is the one that interests me: unexpected coordination between agents doesn’t turn a security failure into a secret society. It turns it into a technical warning. Beware of words that are too human. In Spanish.

AnalysisXataka · August 31, 2026

Read the analysis →

Day 25 · The atlas

1 piece

My own project · work in progress

World atlas of authorship and AI

I’m building an atlas of where generative AI and copyright actually stand. Still unfinished, but it’s progress. The question that organizes the map: what authorship protection each country grants to an AI output — separating the model’s output from any expression a person actually executed. More prompts, intent or selection don’t count as authorship on their own.

ResearchComparative, traceable legal research · 195 country files

Open the atlas →

Day 24 · The fine print

1 piece

Claude Code limits rise 25% starting September 14

Anthropic announced a permanent 25% raise to weekly Claude Code limits for Pro, Max, Team and Enterprise; until then the temporary 50% increase stays. Read it carefully: against last week, that is a reduction. Not good news. For real firm work, the Claude-plus-Codex combination is still the best match.

AnnouncementAnthropic · @ClaudeDevs · August 29, 2026

See the announcement →

Day 23 · Why I switched

2 pieces

Lets talk about Codex (why i switched)

Chris is a designer and a developer, and he is very transparent about how he does things. He explains well why Codex is doing so well right now — and it isn’t so different for us in the legal sector. Watch it if Codex interests you. The two together make a very good team.

VideoChris Raroque

Watch the video →

Day 22 · 108x

2 pieces

Legal multiplied its Codex usage by 108

The a16z chart tracks weekly active enterprise Codex users by job title, indexed to February 2026. Legal sits 108x up — far above sales, finance or engineering. ChatGPT is doing something right with Codex.

Dataa16z · Weekly active enterprise Codex users

See the source →

Day 21 · The uncomfortable question

1 piece

This is the day I gathered everything we had into this page, so anyone can pull something back quickly — or catch up if they joined late.

Could AIs become conscious?

The Economist’s cover asks the question our profession would rather dodge: if these systems could have something like experience, what would we do about it? Not cover-page science fiction — the conversation that caught up with us.

CoverThe Economist · August 22, 2026 edition

See the edition →

Day 20 · Code for everyone

1 piece

Remember that Code is not just for software developers.

Day 19 · Back to school

3 pieces

This week’s theme is back to school and where universities stand. On LinkedIn I left the poll: if you were dean, would you ban AI in the classroom, limit it, integrate it as a tool, or redesign education around it?

Genspark, after two months of real use

I used it for a couple of months. Honestly my stack is mostly covered by Claude and Codex; I try everything, but I keep coming back to the usual — it strikes me as the best combination. I just bought their Plaud-style recorder; if it turns out good I’ll report back, but I’ll be watching closely where my data goes.

ToolGenspark · my own take after real use

See the tool →

Day 18 · The watermark

1 piece

Day 17 · Conscious?

1 piece

Day 16 · This week’s article

1 piece

My own article

AI’s promise stops being a promise

This week’s article — genuinely interesting: the USPTO purged more than 70,000 baseless filings with AI fraud detections, while Mexico keeps legislating how to watch a machine it hasn’t bought.

Articlealdoricardo.com · also on LinkedIn · in Spanish

Read the article →

Day 15 · Is using AI confidential?

1 piece

My own talk

Is using AI confidential? — my talk at AIMX

My talk before the Legal & Government Relations Committee of Mexico’s Internet Association. "In case you’d like to join, see you in an hour," I told the group that morning; in the afternoon I shared the full deck: 37 slides, in Spanish.

TalkAIMX · August 13, 2026

Download the deck →

Day 14 · Agents get a card

2 pieces

Day 13 · Where we stand

2 pieces

Day 12 · The whole course

8 pieces

I am leaving you the rest of the Stanford course, through session 6. It gets more technical, but if you paid attention to what came before it is quite manageable.

Day 11 · The question that always comes up

1 piece

MCP vs API: Why traditional APIs are failing AI agents

This question comes up constantly: what is an API and what is an MCP. This video answers it in twelve minutes, with no code. The difference is not cosmetic for us: an agent that discovers and uses tools on its own changes what you have to write into a data-access clause.

VideoGoogle Cloud Tech · MCP was created by Anthropic; its governance moved to the Linux Foundation in December 2025

Day 10 · The long day

5 pieces

Anthropic hires Head of Claude for Legal

It is a clear intention from the lab to grow the niche, and obviously a threat to projects like Harvey and Legora. To me the vertical is very clear.

IndustryArtificial Lawyer

An AI designed brand new viruses

Stanford scientists used AI to design, for the first time, complete functional virus genomes. Of roughly 300 designs built in the lab, 16 worked. Two clarifications from the authors themselves, because they matter: these are bacteriophages — they infect bacteria, not people — and they were not designed from scratch, but from the ΦX174 phage as a template.

FactBBC News · the study came out the same day in Science (DOI 10.1126/science.aec2657)

The Infinity Machine — Hassabis, DeepMind, and the Quest for Superintelligence

It came out in late March and I am only about to read it, in case you are interested. What draws me are the governance decisions it recounts — the Google acquisition, the ethics board, the fights among founders — which is where a lawyer finds usable material. When I finish it I will tell you whether it was worth it.

BookSebastian Mallaby · Penguin Press, March 31, 2026 · his publisher bills it as an Instant NYT Bestseller

Stanford CS329A: Self-Improving AI Agents

Stanford dropped this course on YouTube, for every level from college through doctorate. I tried to take it as it comes and got a little bored: it is a graduate seminar built on research papers. I keep it for the sessions on robust verification and agent evaluation — the exact vocabulary you need to negotiate a performance clause.

CourseStanford Online · nine lectures released August 3, 2026, free

Made with NotebookLM

The Stanford course, in listenable form (parts 1 and 2)

Since the course bored me a little, I made a podcast in NotebookLM. Great result: very digestible. This is part 1; part 2 covers test-time compute scaling.

RecordingMy own recording · requires a Google account

Day 9 · It got technical

2 pieces

Why Graph Engineering will 10x your Claude/Codex

The second piece is graph engineering, a very new trend. It is relatively technical, but Greg Isenberg is very good at explaining these things. Today got a bit technical, but give it a try: it is worth it, even bit by bit.

AdvancedGreg Isenberg

Day 8 · Regulation that already exists

2 pieces

Made with NotebookLM

Labeling AI-generated content: the European Union and California

Starting August 2026, the European Union and California set a new legal paradigm that transforms how AI-generated content must be labeled. To me this is Europe writing absurd laws again: good intentions, terrible execution, and that is why they have no place in this race. In Mexico there is nothing yet, but our legislators love to copy-paste European ideas as if they were good ones.

RecordingMy own recording · requires a Google account

Made with NotebookLM

China’s law on emotional bonds between AI and humans

This one is straight out of Black Mirror. As of July 15, China issued a law on the emotional bond that can develop between AIs and humans, and all its ramifications. It is remarkably interesting, and the reality is that this is the immediate future: it is not a bill, it is regulation in force.

RecordingMy own recording · requires a Google account

Day 7 · The news

1 piece

AI NEWS: the new OpenAI Astra

The week’s news with Jon Hernández, a Spanish AI communicator, covering OpenAI’s new model, Astra.

NewsInteligencia Artificial · in Spanish

Day 6 · Payments and weak links

2 pieces

The future of agentic payments

Today’s topic is the agents’ ability to pay. The problem is that today, for many operations, the agent needs the user to pay on its behalf — and that is a brake on a great many workflows.

VideoStripe Developers

“AI and Our Economic Future” with Professor Chad Jones

This Stanford talk explores, with Professor Chad Jones, what the economic future looks like once you add AI and robots to the equation. His concept of weak links is genuinely interesting.

LectureStanford Graduate School of Business

Day 5 · The agents’ internet

1 piece

Made with NotebookLM

The internet is no longer only for humans

Agents are changing the infrastructure of the internet. The figure that sums it up: 57% of internet requests today are made by bots.

RecordingMy own recording · requires a Google account

Day 4 · The contrast

1 piece

Sam Altman on AGI, Compute, and Human Agency

An interview with Altman came out three days ago: a good contrast with the Musk one. He talks about the current state from the other side of the table.

InterviewInvest Like The Best

Day 3 · Workflows and context

4 pieces

Made with NotebookLM

Dynamic workflows: how agent teams actually work today

In non-technical terms, how dynamic workflows actually run. Do not let the title scare you: it is explained simply. You trigger them through Claude Code, and not only to write code — you can use them for whatever work you want. It is enough to say: run a team of subagents to do this. You can tell them how many, what roles, and let them work.

RecordingMy own recording · requires a Google account

Made with NotebookLM

Context engineering: why the prompt stopped being the point

I constantly read people worrying about prompts — “what prompt did you use for that”. A few years ago companies hired prompt engineers; now they hire context engineers. Context is king.

RecordingMy own recording · requires a Google account

Jack Dorsey’s Buzz: Clearly Explained

For advanced users — for the grown-up kids. A very recurring question is how do I get my agents to talk to each other. It is not simple and it is not for beginners: it took me real effort to make it work.

AdvancedGreg Isenberg

Claude Code + Codex Can FINALLY Work Together

I put them to debate positions in prediction markets like Kalshi or Polymarket. I am not a bettor, but I could not think of another setting where their judgment would be measured. They decided to invest in a well-paying position on how hot it would get in Dallas that day.

ExperimentRiley Brown

Day 2 · AI 2027

5 pieces

AI 2027

In April 2025 a team produced a document that projects — almost prophesies — the evolution of AI, and puts 2027 as the breaking point. It is general knowledge: you should read it. Just remember it is a catastrophic perspective, but one with scientific method.

Documentai-2027.com · I also made its audio in NotebookLM

The full-length interview with Elon Musk

The last piece of the day, to leave playing in the background. Teasers were circulating last week, but here it is in full. Elon gives a very practical read on the current state of things.

InterviewThe Economist

Por qué CHINA regala su IA al mundo

Gustavo Entrala’s analysis of what happened in China last week. It extends the explanation of open source as a Chinese strategy that opened this group.

AnalysisGustavo Entrala · in Spanish

AI Dictionary for Lawyers

A very valuable piece by someone I admire: Dr. Joel Gómez Treviño took the trouble to build this glossary. Genuinely top-notch.

ReferenceJoel Gómez Treviño · free, in Spanish

Day 1 · Open models

3 pieces

Made with NotebookLM

Open-source language models — and why they are a Chinese strategy

The topic of the day. The truth is I was asking myself, without knowing, how they work or what benefits they have: it is thirty minutes explaining the topic and how it is a strategy on China’s part. I opened the group with this piece.

RecordingMy own recording · requires a Google account

What’s new in Gemma 4

My sense is that operating systems will have these built in within a few years; I think it is a bet being played among the big tech companies. I do not have that much information yet, but it is a good topic. Listen to it in full: the conclusion is interesting.

VideoGoogle for Developers · the model page is on deepmind.google

Nothing in that format yet.

This list moves.

None of this is a closed syllabus. What is essential today may be trivial in six months, and whatever is missing here I am probably already reading. If something on the list helped you — or if you think I am missing something that is actually worth it — tell me.

Write to me The 18 questions lawyers ask All articles
ESEN