Axel Díaz

Desarrollador web y programador de aplicaciones. Entusiasta. No muy cuerdo. Amante de la buena música.

[Axel Díaz @AxelioDiaz]

Theo timed Gemini 4 Carbon: 11 seconds before it starts responding 🐢

Before trusting a benchmark (or a tweet), measure two numbers: TTFT (how long your user stares at the spinner) and total time. time.perf_counter() with stream=True is enough. https://t.co/RBJ6sDa4qX

[Theo - t3.gg @theo]

I dug more into Gemini’s speed issues. Almost all of it is server side. Their inference is awful.

It takes 11 seconds before it starts to respond, 20x longer than Fable.

[Axel Díaz @AxelioDiaz]

DeepMind says AlphaProof Code proved Signal's protocol secure 🔐

For us mere mortals, property-based testing sits between a test and a proof. hypothesis tries 100 weird inputs and finds bugs your "hello" test never will 😅 https://t.co/ONeUGB9MqT

[Google DeepMind @GoogleDeepMind]

For the first time, AI has mathematically proven @signalapp’s protocol is secure down to the code level.

Using our breakthrough AlphaProof Code framework, we verified the messaging app’s cryptographic system – which protects billions of private messages every day.

[Axel Díaz @AxelioDiaz]

ThePrimeagen on vibe coding: every change now takes 3 hours 😂

What keeps me out of that hole:
- a test before the change
- small diffs (git diff --stat)
- ruff check . && pytest -x
- one commit per thing that works

You decide what stays. https://t.co/c5x1gfWojb

[ThePrimeagen @ThePrimeagen]

So I have been vibing

Going hard

Now every change takes 3 hours.. wtf this sucks

[Axel Díaz @AxelioDiaz]

A 100B-parameter model on a laptop with 29 GB of RAM? 🤯 (via @svpino)

Napkin math, params x bits / 8:
16 bits: 200 GB
8 bits: 100 GB
4 bits: 50 GB

29 GB is ~2.3 bits per parameter. Ask how many bits first 😉 https://t.co/ydrhWX5vOk

[Santiago @svpino]

A new open-weight 100B-parameter model that runs on your laptop.

It's fast and good enough for vibe coding.

They shared a few metrics that look pretty impressive:

• Only needs 29GB of RAM, which is good for a 100B model
• Beats Claude Opus 4.6 on Terminal-Bench 2.1

[Axel Díaz @AxelioDiaz]

Google now has Gemini 4 Argon, Carbon and Hydrogen 🧪 I just wanted small, medium and large.

Tip: never hardcode the model name. MODEL = os.getenv("LLM_FAST_MODEL") and you change one env var, not 40 files.

Carbon and Argon go 2x in price on Dec 1. https://t.co/L729f8lye4

[Simon Willison @simonw]

The pricing for Gemini 4 is pretty confusing - Argon and Carbon are the same token price as each other, and the new Hydrogen is only a 25% discount on that price - though Carbon/Argon will 2x in price on Dec 1st

Hydrogen's pelicans are notably worse
https://t.co/XZqY5eQxcD

[Axel Díaz @AxelioDiaz]

@MoureDev just hit 1 MILLION on YouTube and celebrated with a for loop 🤘

for _ in range(1_000_000):
print("gracias")

That 1_000_000 has been valid Python since 3.6. Congrats, Brais 😄 https://t.co/Lwf0UjzmhO

[Brais Moure @MoureDev]

Casi 6 años entre estas 2 placas, pero... Ya somos 1 MILLÓN en YouTube!! 🤘

for _ in range(1_000_000):
print("gracias")

[Axel Díaz @AxelioDiaz]

Gemini 4 Hydrogen costs $1.50 per million input tokens and $7.50 per million output (via @TheRundownAI).

2M in + 500k out = $6.75.

Output costs 5x input, so asking for "short answers" isn't just style, it's savings 💸 https://t.co/U3YRAExPe7

[The Rundown AI @TheRundownAI]

LAUNCHED: Google DeepMind released Gemini 4 Hydrogen, a speed-first model that edges out its own frontier model, Gemini 4 Argon, 51 to 50 on Artificial Analysis' Intelligence Index.

Pricing: $1.50 in, $7.50 out / million tokens.

The rollout starts today across Search, the Gemini app and Google’s developer tools.

[Axel Díaz @AxelioDiaz]

Your CLAUDE.md is not your diary 📓

@realpython nailed it: a longer CLAUDE.md can make Claude Code do worse work. Every line should stop a mistake you've already seen.

Fewer instructions, better answers. Just like meetings 😅 https://t.co/1SQOvVAFJr

[Real Python @realpython]

A longer CLAUDE.md can make Claude Code do worse work. Every line should stop a mistake you've seen.

1. The AWS Bedrock slip one line fixed

2. Why to trim /init's draft

3. Who wins when two files disagree

4. What importing AGENTS.md costs

5. The test for every line

🧵

[Axel Díaz @AxelioDiaz]

Almost 24 hours of Microsoft 365 downtime: complex systems fail suddenly (@emollick).

Python defense: a circuit breaker. After N failures in a row, stop hammering it for a while 🔌

It won't fix the outage, but your app won't join the fire 🔥 https://t.co/gHfv0fkPnA

[Ethan Mollick @emollick]

Complex systems fail suddenly and catastrophically for complex reasons.

[Axel Díaz @AxelioDiaz]

Anthropic announced its IPO and says it has filed a registration statement with the SEC 📈

Serious question: does paying for a Claude subscription count as angel investing? 😅

(Not financial advice.) https://t.co/nkeMlCdXDM

[Anthropic @AnthropicAI]

Our initial public offering: https://t.co/Ah4BkI2loD

Legal disclaimer: Anthropic has filed a registration statement (including a prospectus) with the SEC for the offering to which this communication relates. Before you invest, you should read the prospectus in that registration statement (available at: https://t.co/Z44axxzf9b) and other documents Anthropic has filed with the SEC for more complete information about Anthropic and this offering. You may get these documents for free by visiting EDGAR on the SEC Web site at https://t.co/94ojsWZTKi. Alternatively, copies of the preliminary prospectus may be obtained from: Goldman Sachs & Co. LLC, Attention: Prospectus Department, 200 West Street, New York, NY 10282, by telephone at [phone], or by email at [email protected]; or Morgan Stanley & Co. LLC, Attention: Prospectus Department, 180 Varick Street, 2nd Floor, New York, NY 10014.

The securities may not be sold, nor may offers to buy be accepted, prior to the time the registration statement becomes effective. This post is not an offer to sell, or a solicitation of an offer to buy, the securities in any jurisdiction where the offer or sale is not permitted.

[Axel Díaz @AxelioDiaz]

Anthropic launched Claude Meetings: Claude joins the call and builds docs and whiteboards while the team talks 🎥

Finally, someone in the meeting who will actually read the agenda 😅

Would you let it into your standup? https://t.co/5X0lzROtrN

[Claude @claudeai]

Introducing Claude Meetings: add Claude to your team’s video call and it builds docs, whiteboards, and other artifacts with you while you’re still talking.

Starting to roll out in beta on Team and Enterprise plans.

[Axel Díaz @AxelioDiaz]

B, about 3 seconds. time.sleep(1) blocks the event loop, so the three tasks run one after another. Swap it for await asyncio.sleep(1) and they can wait together: about 1 second. async def doesn't make blocking code non-blocking 🙂

[Axel Díaz @AxelioDiaz]

Quick Python quiz 🐍 About how many seconds does this take?

A) 1
B) 3
C) Depends how much faith you have in asyncio 🙏

Answer in the replies later. No running it first :)

[Axel Díaz @AxelioDiaz]

New post: FastAPI tests that don't rot. Swap dependencies, roll back every test, and the TestClient lifespan gotcha. Tested on FastAPI 0.162. https://t.co/2jWhXttDk7

[Axel Díaz @AxelioDiaz]

New post: Django migrations that don't take your site down. db_default, batched backfills, the 3-step NOT NULL, and why you never rename a column in one deploy. Tested on Django 5.2. https://t.co/lEQGF1Gs7D

[Axel Díaz @AxelioDiaz]

The model can sketch the furniture. Python can make the shopping list 🪵 A few rows and csv.writer are enough to export parts, quantities and materials. I'd still check the measurements before buying wood. Or sitting on the result 😅

[@levelsio @levelsio]

Claude Code can design furniture part by part and then give you a bill of materials to order each part and put it together yourself

And way cheaper than big interior brands sell you their furniture

Furnishing my house I discovered how absolutely insane prices are for the most basic cheap materials put together with a fancy brand tied to it

One shop had a sofa for $30,000 😂

I think with Opus 5.5 and other models easily operating 3d programs like Blender and FreeCAD (both free) it makes more sense to cut out the middle man and go straight to suppliers and design and build your own furniture

Which in a way makes customers more powerful (they can get exactly what they want and usually for cheaper without the markup)

Anyway @theannalux designed this table from scratch, she's not a furniture designer, not a coder, she just had an idea of a style and worked with Claude Code in the terminal (far superior to the app I will keep saying this)

If you design your own furniture and buy the raw materials directly, you can save 85% compared to premium brands like Poliform

That means the $30,000 sofa becomes $4,500! And the $8,000 dinner table becomes just $1,200!

You can see where the idea comes from that AI will make everything 5x to 20x cheaper, and it's already starting!

[Axel Díaz @AxelioDiaz]

An agent calling an API doesn't get a free pass on security 😅 I still want secrets out of source, untrusted input checked, and the smallest permissions the job needs. Especially when the input is something another agent found.

[Greg Brockman @gdb]

security architecture fundamentals are more important than ever before

[Axel Díaz @AxelioDiaz]

If I need one API endpoint, I don't also need a React app, a monorepo and a weekend 😅 FastAPI can start with: from fastapi import FastAPI; app = FastAPI(). Add a route, then add complexity when there's an actual reason.

[Miguel Ángel Durán @midudev]

Lo peor de la IA es esto. Siempre el mismo stack por defecto: Next.js, React, Tailwind y PostgreSQL.

Y son tecnologías que me gustan, eh. Pero me da pena pensar que se van a perpetuar y será difícil que nazcan nuevas y mejores alternativas en el futuro.

[Axel Díaz @AxelioDiaz]

Before indexing 47,000 videos, a tiny Python quiz 😅 What does this print? A) Error B) ['a', 'b', 'c'] C) ['a', 'b', 'a', 'c']. Answer in the replies later. No running it first :)

[Andrej Karpathy @karpathy]

Earlier this week Tristan published his top YouTube channels:
https://t.co/NZ3UCNo6cT
Which looked excellent so clearly I had to turn it into TristanTube:
https://t.co/l0KyW1WdNH
47K super nerdy videos +TV mode for laid back watching. (One-shotted with Claude Code in 3 minutes)

[Axel Díaz @AxelioDiaz]

More tokens doesn't have to mean a bigger bill. Ramp's late-summer data, reported by TechCrunch, is a nice reminder. I still want budgets, smaller models where they work, and less irrelevant context. The goal is useful work, not a token high score 😅

[TechCrunch @TechCrunch]

AI usage keeps climbing, but spending has stalled. Price cuts and smarter cost controls among the biggest spenders could challenge investor assumptions about AI hypergrowth.
https://t.co/unPWCeaKGW

[Axel Díaz @AxelioDiaz]

DeepMind's RSA-1024 research isn't 'all encryption is broken'. The results target legacy keys; modern RSA-2048+ isn't threatened by these attacks. Here's the toy arithmetic in Python. Tiny primes, no padding, absolutely not production crypto 😅

[Google DeepMind @GoogleDeepMind]

We’re sharing new research demonstrating how AI agents can accelerate the science of stress-testing encryption.

Guided by humans, our models uncovered two major weaknesses in legacy 1024-bit RSA cryptosystems.

Our VP Science @Pushmeet explains more ↓ https://t.co/SqNkBbkih2