Matthew Evans

Aspiring Tier-1 ML Operator | Sim-to-Real | Reinforcement Learning | Isaac Sim+Lab | Local Inferrence

[vittorio @IterIntellectus]

sorry to blackpill but this is genuinely terrible news

the whole “AI is going to kill everyone” panic is becoming a massive regulatory capture attempt by the largest AI labs (exactly what you’d expect)

this would only slow down american AI while everyone else keeps building

you can't stop AI. the only question is who gets there first

[Matthew Evans @MatthewEvans87]

Never forget.

[LAURENCE HAIM @lauhaim]

Pour tous ceux qui ont vécu de près le 11-Septembre, il reste un débat extrêmement douloureux : celui des « jumpers », ces hommes et ces femmes qui, prisonniers dans les tours, ont choisi de sauter dans le vide plutôt que d’affronter les flammes et la fumée.

On estime qu’ils furent plus de 200.

Lorsque les premières chutes ont commencé, les directeurs et stars des télévisions américaines ont été confrontées à une question terrible, en plein direct : doit on montrer ces images ?

Certaines les ont diffusées dans les premières minutes, avant de décider de ne plus les montrer . Le debat a ete intense .

Il ne s’agissait plus seulement de raconter ce qui se passait, mais de décider, en quelques secondes, jusqu’où pouvait aller le devoir d’informer face à la mort d’êtres humains filmée en temps réel.

Vingt-cinq ans plus tard, le débat revient sous une forme totalement différente. Sur les réseaux sociaux, ces images circulent de nouveau. Certains estiment qu’une génération qui n’a pas connu le 11-Septembre doit les voir pour comprendre réellement l’horreur de cette journée.

Aux États-Unis, la question des « jumpers » reste toujours l’un des sujets les plus sensibles de la mémoire du 11-Septembre.

Et derrière ces images demeure une question journalistique qui, elle, n’a pas vieilli : que doit-on montrer en direct lorsqu’une tragédie se déroule sous nos yeux ? Où s’arrête le devoir de témoigner et où commence le devoir de protéger la dignité de ceux qui meurent ?

[Matthew Evans @MatthewEvans87]

qwen3.8 + ollama + copilot cli + RTX5090 = Win.

I've been using this combo in a mature codebase for a week or so. It's incredible how far local inference has come in the last 6 months. My hardware hasn't changed, the models have.

[Matthew Evans @MatthewEvans87]

@XFreeze 😈

[Matthew Evans @MatthewEvans87]

@USTechWorkers 😳

[SightBringer @_The_Prophet__]

⚡️Autonomous driving looks like a transportation upgrade, but economically it is a reorganization of mobility itself.

The core change is simple:

The car becomes productive capital.

Today, most cars are expensive assets that sit idle and depreciate. Once the driver disappears, the vehicle can operate continuously, serve many users, generate cash flow, and be financed like infrastructure.

That shifts value from car ownership to network ownership.

The real prize becomes owning the autonomy stack, fleet, charging, dispatch, customer relationship, and ultimately the economics of every mile traveled.

The endgame is owning the mile.

Labor takes the first hit. Taxi drivers, rideshare drivers, delivery drivers, chauffeurs, and eventually parts of trucking lose the scarcity premium attached to human driving. Transportation gets cheaper because the human operator disappears from marginal cost.

Cheaper mobility probably creates more mobility. People may own fewer cars while total miles traveled rise. Children, elderly people, non-drivers, commuters, and businesses all gain access to cheaper movement.

Then geography starts bending.

Parking loses value. Urban land gets released. At the same time, commuting distance becomes less painful because travel time becomes usable time. That expands the viable radius around cities and changes housing, labor markets, and commercial real estate.

Short-haul aviation, rental cars, buses, and some rail also come under pressure. Door-to-door autonomous travel removes enormous friction.

Insurance changes too. If autonomy becomes materially safer than humans, the entire accident economy contracts: claims, repairs, towing, litigation, rental replacement. Liability migrates toward software, manufacturers, and fleet operators.

And then comes the hidden layer:

driving time becomes monetizable attention.

The cabin becomes a screen, office, game room, shopping surface, entertainment platform. The company controlling the vehicle controls another slice of human attention.

So the full economic transition is:

cars become yield-producing assets, fleets replace some ownership, software captures more value, driving labor gets repriced, land gets repriced, travel expands, accident economics shrink, and mobility becomes a platform business.

The macro consequence is the same pattern AI is creating elsewhere:

productivity rises, prices fall, output expands, labor becomes less necessary, and capital ownership becomes more valuable.

That is the real fracture line.

Once the machine drives itself, the central economic question becomes:

Who owns the machine, the software, and the network?

Because whoever owns that layer can collect rent from motion itself.

[Matthew Evans @MatthewEvans87]

Is it just me or does Sonnet 5 delivery terrible performance for more money? I haven't been this frustrated since GPT 3.5.

[Matthew Evans @MatthewEvans87]

@karenxcheng @cursor_ai I love this.
It’s totally impractical, and yet, I love it.

[Matthew Evans @MatthewEvans87]

Built a real-time two-threaded frame processing pipeline in C++.

Camera feed → ring buffer → transform chain → live display. Grayscale, Gaussian blur, Sobel edge detection, threshold — running on live frames at camera rate.

The interesting part isn't the image processing. It's the seam between the threads.

A fixed-capacity, heap-free ring buffer decouples the capture thread from the process thread. Mutex and condition variable at the boundary. Overwrite-on-full policy — in a real-time pipeline, a stale frame is worse than a dropped one, so the producer is never blocked. Atomic flag for clean shutdown.

Things that bit me:
→ cv::Mat stride padding corrupting frame data on copy
→ BGR vs RGB silently producing wrong luminance weights
→ NVIDIA driver noise in Valgrind masking clean heap output

Valgrind on clean exit: zero leaks, zero errors in my code.

Project 3 of a C++ learning arc targeting perception and autonomy engineering.

https://t.co/04TZeRdPxV

[Matthew Evans @MatthewEvans87]

Spent a few hours this week loading my SO-101 manipulation policy into a C++ binary via LibTorch.

Encoder forward pass, observation assembly, policy forward pass. GoogleTest suite confirming C++ output matches Python reference to 1e-5.

Not a production system. But I now know the handoff: TorchScript is how a trained model crosses from Python into a C++ target, and numerical parity testing is how you verify it didn't break in transit.

That handoff is most of what "deploying ML" actually means in a real system.

https://t.co/rDCnY9ZniZ

[Matthew Evans @MatthewEvans87]

Check your data before consumption:

Fun finding while exploring the Expedia hotel search dataset this week.
Positions 5, 11, 17, 23 have ~5% of the click volume AND exposure of neighboring positions. Evenly spaced at intervals of 6. Pattern stops at position 29.
Not low relevance -> low exposure. Consistent with a non-organic display slot type interleaved at irregular fill rates. i.e., Ad slots.

[Matthew Evans @MatthewEvans87]

1. This will be resisted strongly by the establishment
2. This is the beginning of an alternate healthcare system akin crypto adjacent to fiat

[Midjourney @midjourney]

A technical dive inside our new "Midjourney Scanner"

[Matthew Evans @MatthewEvans87]

The beginning of the end. It was nice while it lasted, but like the internet of yesteryear, all good things must come to an end.

[Anthropic @AnthropicAI]

The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.

The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance.

Access to all other Claude models is not affected.

We apologize for this disruption to our customers. We believe this is a misunderstanding and are working to restore access as soon as possible.

Read our full statement: https://t.co/bwn0sximKZ

[Matthew Evans @MatthewEvans87]

SO-101 Sim2real efforts continue. Improvements to camera tuning, DR, and action clamping + EMA smoothing.

[Matthew Evans @MatthewEvans87]

Working to improve sim-to-real policy transfer on so101.
Added EMA and action clamping (excluding gripper) in training ➡️ now the policy has trouble grasping the object ☠️

[Matthew Evans @MatthewEvans87]

@mattvanswol You don’t understand the GOP but you (must) understand politicians. You (must) understand people.
It’s all a game when it’s other people’s PROPERTY (life, money, etc.) on the line.

[Shawn Ryan @ShawnRyan762]

$32 million dollars and an entire administration mobilized to destroy one congressman.
His crime?
Demanding answers about Epstein class abuse networks, and refusing to let child predators hide behind political cover.
If that level of firepower doesn't tell you who's being protected, nothing will.
Go Massie!

[Matthew Evans @MatthewEvans87]

@PalmerLuckey Wait, Are you building an Arwing?
Does it do a barrel roll?
I assume it has <B> bombs?

[Matthew Evans @MatthewEvans87]

@PalmerLuckey Don’t blow up that Arwing!

[Elon Musk @elonmusk]

Beware the empathy exploit.

Empathy is good and right when thought through (deep), but can be deadly to civilization when simply stimulus-response (shallow).

For example, releasing a repeat violent offender may feel good at first (shallow empathy for the criminal), but it is wrong to do so when that person will go on to hurt or murder innocent victims, as there should be deep empathy for future victims.