Vignesh

[Thinking Machines @thinkymachines]

Today, we are introducing Inkling.

Inkling reasons efficiently across text, image, and audio modalities. We are making the full weights available.

https://t.co/Ghebq5mG30

Available today for fine-tuning on Tinker. Play with it in the Inkling Playground. 🧵

[ΛI DRIVR @AIDRIVR]

inhuman reactions

[Xiaopeng He @xiaopenghexpeng]

Ahead of the big brand day, I'm pleased to share that following its successful rollout in China this March, XPENG's Next Generation Pilot (NGP) intelligent driving system powered by VLA 2.0 technology will launch in Europe in early 2027. We've been testing on European roads for quite some time and results are pretty positive as in the video. I will also be testing it on the Munich road later today. Your turn @liuxianming !

And we're not stopping there. We look forward to working with European partners to bring XPENG Robotaxi fleet powered by the same foundation model to local markets. Together, we'll accelerate the future of autonomous mobility.

[Xianming Liu @liuxianming]

We've been driving the streets of Munich, validating every detail over the past few days.

More information will be revealed at today's event. See you at 14:00(GMT +2) today!

[Wei ZHAN @Wei_ZHAN_]

To obtain robust autonomy, it has always been the desire for planners to experience large-scale, high-value scenarios in ultrafast closed-loop training with minimal unit cost of data and compute, ideally self-evolving even without offline human demonstration.

TerraZero is our straightforward response at @AppliedInt to such a desire by scaling self-play reinforcement learning with procedural simulation and zero demonstration:
- Insanely fast in closed loop, i.e., 2.8 million simulation steps per second on 8 GPUs;
- Endlessly streaming hard and diversified corner cases with a procedural generation;
- Robust self-play recipe trading sample efficiency for compute efficiency;
- Zero demonstration in training, and zero-shot emergence across domains (cities, datasets, etc.);
- Minimal unit cost for experiencing high-value data points with a highly scalable framework.

So excited to see that one stack powered by TerraZero can serve various outcomes:
- Setting the state-of-the-art in the challenging, long-tail planning benchmark (InterPlan) with a clear edge;
- Topping the realism among demonstration-free methods in sim agent benchmark (WOSAC);
- Comprehensive practicality with policy covering heterogeneous agents (VRUs, trucks, trailers, etc.) and tackling noisy policy inputs.
On our website, you can also select arbitrary combinations of scenario/agent/view features and see how TerraZero works - play with it!

Really impressed by the potential of large-scale, high-value synthetic data in closed loop powered by ultrafast training, and TerraZero is just the beginning of our Terra Series research - committed to pushing the boundaries of physical AI learned in closed loop with scaled simulation and minimal demonstration. More follow-up work will be announced soon, stay tuned!

TerraZero report: https://t.co/iJA0NdHkF8
TerraZero website: https://t.co/aaqJpNCduM
Terra Series website: https://t.co/5s4CiOoBd5

[Applied Intuition @AppliedInt]

"The upside is there's more opportunity."

Carolyn Davis, U.S. Automotive Lead here at Applied Intuition, gives her take on driving impact at a high growth company.

https://t.co/B0CK0OOJ8w

[Xianming Liu @liuxianming]

https://t.co/z5SWfoUnSV

[Applied Intuition @AppliedInt]

We're building physical AI for every moving machine.

🎧 Tune into the full @latentspacepod episode: https://t.co/eijgTwpzD4

[Tesla @Tesla]

Engineering tests of the first production Cybercab have begun in Austin

[Mike P @mikepat711]

Amazing V14.3.4 clip that @jtesla47 just sent me and asked me to share from his M3P.

What I love most about this clip is that FSD understands what needs to happen and reacts BEFORE the lead car even throws it in reverse. This is such a critical indicator that FSD isn’t just reacting to objects in its path, but actually understands what happens next before it even needs to start backing up to clear the dodge ram.

This is one of my favorite edge cases in a minute.

FSD IS THE TRUTH

[Applied Intuition @AppliedInt]

本日、Applied Intuitionの自動運転システムが正式に日本へ上陸しました。日本の自動車業界は、サプライヤーに対して独自の基準を課しています。

私たちは長年にわたり、日本に拠点を持つチーム、日本でのデータインフラ、日本の企業とのパートナーシップを構築し、その基準を満たす準備を進めてきました。

自動運転を世界中でスケーラブルにするためのこの重要な一歩について、詳しくはこちら：https://t.co/mJPTg3i2GS

Today, Applied Intuition's Self-Driving System officially comes to Japan.

The Japanese automotive industry holds its suppliers to a different standard.

We've been building to meet it for years: local teams, local data infrastructure, local partnerships.

Learn more about this important step toward making autonomy scalable worldwide: https://t.co/4neCmik5nk

[Elon Musk @elonmusk]

Upcoming releases of FSD will remember your parking preferences, so that the car goes to the right location at your home, office, school drop off, etc.

Destination parking is by far the biggest reason people now intervene with FSD. Critical safety interventions are extremely rare.

[Vignesh @vignesh]

Some good stuff here. Pure vla doesn’t fit autonomy.

[Fred Lambert @FredLambert]

I had a fun interview with Dr. Xianming Liu, head of AI at @XPENG_Global, to gain some insights into the company's progress toward autonomous driving and physical AI.

It's impressive what they have been able to do in a short period of time. VLA 2.0 is now roughly at parity with Tesla FSD.

[Vignesh @vignesh]

Congrats on CVPR 2026 Best Paper 🏆 Amazing to see this on the humble single encoder transformer architecture. Super smart to avoid per frame decoding.

[Chuhan Zhang @ChuhanZhang5]

A SINGLE encoder + decoder for all the 4D tasks!
We release 🎯 D4RT (Dynamic 4D Reconstruction and Tracking).

📍 A simple, unified interface for 3D tracking, depth, and pose
🌟 SOTA results on 4D reconstruction & tracking
🚀 Up to 100x faster pose estimation than prior works

[BigHu-L5 @CyberTshr]

This is absolutely stunning. With FSD, your attention is freed from the steering wheel, giving you more time to enjoy the journey and the beauty around you.
@Tesla @Tesla_AI @elonmusk @aelluswamy @elonmusk @edgecase411 @wholemars #FSD

[Tesla AI @Tesla_AI]

.@aelluswamy will discuss our approach to foundation models for robotics at CVPR today

Room 603, 3:30pm local time

Covering architecture, large-scale multimodal training, end-to-end control, safety & deployment

Come swing by booth 255 for live demos after!

[Steven Peeters @aikisteve]

You were all asking to not intervene @Tesla FSD when it gets into a pickle. Well, here you have an example where FSD encounters a van on a very narrow🇳🇱road and it squeezes itself in the tiniest gap with only centimetres to spare on each side! 😱 This one definitely puckered up our buttholes 😂 @pvandamcom @teslaeurope @TeslaOwnersNL @SawyerMerritt @wholemars @KRoelandschap #FSDEurope

[Shon 📐 NowTeslafy @nowteslafy]

Here’s a long standing pet peeve of mine for FSD: waiting until all lanes are clear to make the left… this even has a purposeful merge lane.

[Ashok Elluswamy @aelluswamy]

Soon it’ll be driving itself in to Austin city, reporting for duty!

[Dwarkesh Patel @dwarkesh_sp]

Recently met @srush_nlp and he started giving me an impromptu lecture on how targeted on-policy self-distillation works.

I asked him if I could record it on my iPhone.

The basic idea is this: if the model made a mistake at some point in the rollout (for example, calling a tool that doesn't exist), we want to discourage this specific error, but we don't want to just learn from the final reward, because it's a very noisy signal spread out over the whole trajectory.

So we have another model read this trajectory and figure where the error was made. It simply inserts some hint tokens to the part of the trajectory right above where the mistake was made.

Now with these injected hint tokens, have the model run a forward pass. You're not having to regenerate a new rollout - aka no new decode required.

The hint causes the model to assign lower probabilities to the error tokens. You then trains the original model to match these new probabilities, teaching it to downweight that specific mistake.