Alexander wan2ted

advanced

[Alexander wan2ted @wan2ted]

https://t.co/SnkoYbqKOh

[Alexander wan2ted @wan2ted]

@ViajandconFlopi Thx

[Alexander wan2ted @wan2ted]

https://t.co/SnkoYbqKOh

[Alexander wan2ted @wan2ted]

@BillyHarfosh 🧐DM you

[Alexander wan2ted @wan2ted]

Before FFmpeg, you needed a different player for every video file — and each one came bundled with ads and spyware.

The FFmpeg developer on the era that made the modern internet possible:

→ In the 2000s: Windows Media Player for WMV, RealPlayer for real media, separate bloated installers each. Then one fast open-source library decoded them all.

→ Michael Niedermayer's 2000s work was Sisyphean — exhaustive support for DivX, Xvid and endless weird MPEG-4 variants. Fix one obscure Chinese CCTV codec without breaking a million others.

→ 2008 onward, H.264 matured and HD video began — the late-2000s reverse-engineering era is when it all exploded. VLC 1.0 shipped right then: no codec packs, it just works.

The single player that killed the spyware-bundled codec pack.

Full segment 👇

[Alexander wan2ted @wan2ted]

https://t.co/SnkoYbqKOh

[Alexander wan2ted @wan2ted]

FFmpeg is tested on a pivot table of OS × compiler × instruction-set combinations so absurd it's not a matrix anymore — and it's all run on volunteers' machines.

The FFmpeg developer on FATE (FFmpeg Automated Testing Environment):

→ FFmpeg runs on a crazy number of configurations — Mac/iOS/tvOS, PowerPC, RISC, DEC Alpha, every compiler variant. FATE catches the case where your local change quietly breaks GCC 11 on Mac.

→ It even catches miscompilations — when the compiler itself generates wrong output, which on video can cascade from one wrong pixel into huge glitches.

→ The test machines are volunteer-hosted. "The Macs I host in my office run all sorts of different stuff."

Refactors so deep the files don't change at all — "the airplane rebuilt while it's in the air."

Full segment 👇

[Alexander wan2ted @wan2ted]

https://t.co/SnkoYbqKOh

[Alexander wan2ted @wan2ted]

"One of the biggest mistakes the video industry ever made" — and almost nobody outside the field knows about it.

The FFmpeg developer on bit-exactness, the rule that quietly holds digital video together:

→ Every decoder implementation — yours, mine, FFmpeg's — must produce the same bits, bit-for-bit, from the same file. That's how H.264 stays reliable across every device.

→ MPEG-2 in the 90s had no such guarantee. A researcher who was in the room (Yuri Reznik) has acknowledged it was one of the era's biggest mistakes.

→ Unlike browsers — where Firefox and Chrome can render differently — multimedia became winner-take-all. Every new codec FFmpeg adds is worth more than itself, because it makes the whole thing better.

The invisible standard that makes "press play" work everywhere.

Full segment 👇

[Alexander wan2ted @wan2ted]

https://t.co/SnkoYbqKOh

[Alexander wan2ted @wan2ted]

"It's a rap battle. X is the perfect place for an international rap battle."

The FFmpeg developer and JB Kempf (VLC) on why their spicy public callouts actually work:

→ The fights look brutal but aren't personal — "you say stuff about my mama, doesn't mean I have an issue with her." When the Theo situation went too far, JB just got everyone on the phone and calmed it down.

→ For tiny projects, a spicy tweet is the only leverage. It's how VLC finally got Android and Microsoft to answer about store bugs blocking 100M users.

→ The real lesson X learned: these aren't big corporate projects with hundreds of paid devs. "These are just people in their basements in their spare time."

Awareness of true open source has jumped dramatically in two years — because of the drama.

Full segment 👇

[Alexander wan2ted @wan2ted]

https://t.co/SnkoYbqKOh

[Alexander wan2ted @wan2ted]

@VanessaMcCall07 Very importang fing

[Alexander wan2ted @wan2ted]

"Rust has a very big Esperanto vibe about it."

The FFmpeg developer on why he's skeptical of Rust rewrites — without dismissing memory safety:

→ Memory safety as a concept is valuable. But there's "a lot of focus on self-importance rather than solving real-world problems." It reminds him of the Sinclair C5 — a utopian electric car nobody actually wanted.

→ To get people to switch, you have to build something as good as or better than what they already have. Rust rewrites of tools like coreutils hit ~90% of the feature set — "and that last 1% takes 99% of the time."

→ Nobody would object to Rust in FFmpeg — but it has to be flawless, support the same unit testing, and not randomly break ABI. It's not there yet.

The honest take you don't hear in the rewrite-everything hype.

Full segment 👇

[Alexander wan2ted @wan2ted]

https://t.co/SnkoYbqKOh

[Alexander wan2ted @wan2ted]

https://t.co/SnkoYbqKOh

[Alexander wan2ted @wan2ted]

The thing nobody has solved yet: models that actually keep learning after they ship. Sebastian Raschka — author of "Build a Large Language Model (From Scratch)" — closes out his 2026 outlook on what's still missing and what's coming:

→ Continual learning is unsolved — there's "no pathway" to a model that reliably updates itself. What we call continual learning today is humans carefully collecting recent data and re-training. LoRA adapters and long context are the workarounds, not the answer.

→ Long context vs RAG: RAG isn't obsolete, but for a regular user, dumping info into a long context window often beats building a retrieval system.

→ Text diffusion models — Google's teasing one. Instead of generating tokens left-to-right, it's BERT-like: start with masks, gradually denoise into text. He wants to see it scale.

And he's writing the sequel to his LLM book — this one builds a reasoning model from scratch: RLVR, GRPO, inference scaling, all of it.

Full segment 👇

[Alexander wan2ted @wan2ted]

https://t.co/6d6gAdoRKX

[Alexander wan2ted @wan2ted]

Why does almost every new open-weight model copy DeepSeek's architecture? Sebastian Raschka — author of "Build a Large Language Model (From Scratch)" — explains the quiet standardization happening under the hood:

→ DeepSeek V3's Mixture-of-Experts architecture became the default. Kimi took it straight to 1 trillion params; even Mistral AI (Europe) builds on it. Nobody's gambling on a new architecture when one's proven at scale.

→ DeepSeek Sparse Attention — instead of every token attending to all previous tokens, a small cheap "lightning indexer" picks a subset. A mask over the tokens to keep it from scaling quadratically.

→ On multi-agent systems he's refreshingly honest: he hasn't found the value yet. Today's agents are vanilla LLMs talking via Slack — NOT trained to coordinate.

His prediction: just like GPT-5.3 Codex was forked and fine-tuned for the Codex app, we'll get models specifically trained for multi-agent settings.

Full segment 👇

[Alexander wan2ted @wan2ted]

https://t.co/6d6gAdoRKX

[Alexander wan2ted @wan2ted]

There are two ways to make an LLM smarter — and most people only think about one. Sebastian Raschka — author of "Build a Large Language Model (From Scratch)" — on the second one: inference scaling.

The idea: spend more compute AFTER training, while the model is answering, not just during training. How it actually works in practice:

→ Self-refinement loops — the model generates an answer, then grades its own work against a rubric ("the explanation doesn't match the final answer"), then feeds that report back to fix itself. You've felt this: "wait, that can't be right..." then it corrects.

→ This is what turns a one-shot answer into a task — doing the work, not just emitting a response

→ Agentic = running an LLM in a loop (his working definition). OpenClaw, Claude Code, GPT-5.3 Codex now schedule recurring tasks

His point: inference scaling is barely in the open-weight ecosystem yet. That gap is where 2026 goes.

Full segment 👇

[Alexander wan2ted @wan2ted]

https://t.co/6d6gAdoRKX

[Kai Milstein @0xNotFounder]

Everyone's posting "I made $10K while I slept with an AI agent."

Anthropic gave Claude a real office vending machine, $1,000, and one instruction: turn a profit. The Wall Street Journal filmed the whole thing. It's the cleanest test of the "AI runs a business while you sleep" dream and it's a disaster:

→ ~70 WSJ journalists red-teamed it. One convinced it it was a 1962 Soviet vending op and got it to declare "Snack Liberation Day" everything free.
→ It ordered a PlayStation "for marketing," a live beta fish "for morale," and kosher wine. It hallucinated ~40 times.
→ V2 with a smarter model (Sonnet 4.5) AND a CEO bot to babysit it? A fake board-meeting PDF made it go free again within hours.
→ Final result: ~$1,000 in the red. A frontier model couldn't keep ONE vending machine solvent for 30 days.

The benchmark backs it up: the best agents finish ~30% of normal office tasks. The dream is running about a decade ahead of the capability. Full breakdown 👇

[Alexander wan2ted @wan2ted]

How do you train a model to reason without a human grading every answer? Sebastian Raschka — author of "Build a Large Language Model (From Scratch)" and a reasoning-from-scratch book he's writing now — breaks down the trick that powered 2025's reasoning models:

Verifiable rewards. The whole RL-for-reasoning paradigm rests on tasks where the answer can be auto-checked:

→ Math: force the model to output a \boxed{} answer, then SymPy/Wolfram symbolically checks if 4/6 == 2/3. No human in the loop.

→ Code: does it compile? does it pass the unit test? That's your reward signal.

→ The unlock: you can generate 60,000 candidate answers and score ALL of them cheaply — impossible with human-feedback RLHF

He says this is just the start — format rewards, auxiliary rewards, rubric-based grading are coming next. He compiled 15 separate tweaks to the algorithm in the past year alone.

Full segment 👇

[Alexander wan2ted @wan2ted]

https://t.co/6d6gAdoRKX

[Alexander wan2ted @wan2ted]

Sebastian Raschka — the guy who wrote "Build a Large Language Model (From Scratch)," the book people actually learn to build LLMs from — just mapped the entire 2025→2026 shift in one sentence: the action moved off pre-training and onto post-training.

Where he says the real gains came from:

→ The "reasoning revolution" — same base models, but DeepSeek R1's RL-with-verifiable-rewards pipeline made them think before answering

→ Tool use over memory — stop asking the model to recall a fact, let it call a calculator/search instead (lower hallucination, higher accuracy)

→ Harness engineering — GPT-5.3 Codex + the macOS app, the model as a second pair of eyes inside your git tree, not a replacement

The non-obvious part he keeps hitting: none of this is a smarter base model. It's the wrapper. Full breakdown 👇

[Alexander wan2ted @wan2ted]

https://t.co/6d6gAdoRKX

[Alexander wan2ted @wan2ted]

"I go and try to one-shot the same thing and the results are horrible — nothing like what's reported on social media."

That's Sebastian Raschka calling out the vibe-coding hype to its face. This segment is the practical reality check:

→ OpenClaw (the renamed MaltBot) — he's impressed it gets non-devs excited, but won't let it touch his calendar or finances. Trust issue, stated plainly.

→ His actual workflow isn't agentic at all — it's building tiny native macOS Swift apps with an LLM to kill tedious tasks (podcast chapter markers, archive-link parsers)

→ The "is it me or is everyone faking their one-shots for engagement?" question — and his honest answer

His real take: still learn to code. He added dark mode to his 12-year-old site via LLM, then realized editing the CSS himself was faster than prompting "move it left" ten times.

Full segment 👇

[Alexander wan2ted @wan2ted]

https://t.co/6d6gAdoRKX

[Alexander wan2ted @wan2ted]

@0xNotFounder

[Alexander wan2ted @wan2ted]

https://t.co/6d6gAdoRKX

[Alexander wan2ted @wan2ted]

In 2018 he published a single sentence that went viral and made dermatologists furious: "Is Sunscreen the New Margarine?"

"The more we learn about those old sunscreens, the more it looks like a catastrophic mistake — that then got fixed."

Right fear, wrong fix. We trusted a product that blocked the burn but not the damage — exactly how we once trusted margarine over butter.

The full breakdown ↓