Igor Pener (@igorpener) profile picture
Earth
Joined August 2016
Photo by Igor Pener (@igorpener)Photo by Igor Pener (@igorpener)Photo by Igor Pener (@igorpener)Photo by Igor Pener (@igorpener)Photo by Igor Pener (@igorpener)Photo by Igor Pener (@igorpener)Photo by Igor Pener (@igorpener)Photo by Igor Pener (@igorpener)Photo by Igor Pener (@igorpener)
This is actually a bigger deal for humanity than most people can imagine. I will always remember the looks on the engineers’ faces in the Starbase conference room in December 2020 when @elonmusk suggested, then insisted, that the Mechazilla launchpad arms could catch a returning Starship.
Mechazilla catches Starship! Congratulations to the entire SpaceX team on this incredible achievement!
489
2,174
168
20,489
2,590,123
Igor Pener retweeted
How many pencils, how many humans, how many hours would it have taken to recreate the Adobe suite by hand?
Todos los productos de Adobe reimplementados desde cero, gratuitos y de código abierto → getartcraft.com/apps
324
359
31
8,304
590,907
Igor Pener retweeted
Anthropic leaked 512,000 lines of Claude Code source code yesterday. What happened in the next 12 hours is absolutely wild. 4 AM. Anthropic pushes an update to npm. Inside the package: their entire codebase. A 60 MB debugging file accidentally bundled in. 23 minutes later, researcher Chaofan Shou spots it. Downloads the zip. Posts it on X. Within 6 hours: 3 million views. By the time Anthropic’s team woke up, the code was forked 41,000+ times across GitHub. Anthropic started firing DMCA takedowns. Too late. A Korean developer named Sigrid Jin woke up to his phone exploding. He’s Claude Code’s biggest power user. WSJ reported he burned through 25 billion tokens last year. He read the leaked code. Rewrote the entire thing in Python in 8 hours. His repo hit 30,000 stars faster than any GitHub project in history. Then he rewrote it again in Rust. That version now has 49,000 stars. Someone mirrored it to a decentralized platform with one message: “will never be taken down.” The code is permanent. Anthropic cannot get it back. Here’s the part I can’t stop thinking about: Anthropic built something called “Undercover Mode.” Its only job: prevent Claude from accidentally leaking internal secrets. They shipped an entire anti-leak system in their own product. Then leaked their own source code in a .map file. Irony is beautiful
Claude code source code has been leaked via a map file in their npm registry! Code: pub-aea8527898604c1bbb12468b…
244
1,213
173
6,229
1,606,113
Igor Pener retweeted
I (finally) put together a new LLM Architecture Gallery that collects the architecture figures all in one place! sebastianraschka.com/llm-arc…
198
1,426
128
8,081
740,601
Igor Pener retweeted
Everything is Computer, but Computer isn't Everything!
727
1,552
536
12,824
51,219,333
Igor Pener retweeted
Replying to @robustus
Code itself will go away in favor of just making the binary directly. The next step after that is direct, real-time pixel generation by the neural net.
911
617
373
7,287
1,410,339
Igor Pener retweeted
New Engineering blog: We tasked Opus 4.6 using agent teams to build a C compiler. Then we (mostly) walked away. Two weeks later, it worked on the Linux kernel. Here's what it taught us about the future of autonomous software development. Read more: anthropic.com/engineering/bu…
839
2,443
1,369
20,995
8,539,368
We're helping AI to see the 3D world in motion as humans do. 🌐 Enter D4RT: a unified model that turns video into 4D representations faster than previous methods - enabling it to understand space and time. This is how it works 🧵
88
421
86
3,087
370,123
I was inspired by this so I wanted to see if Claude Code can get into my Lutron home automation system. - it found my Lutron controllers on the local wifi network - checked for open ports, connected, got some metadata and identified the devices and their firmware - searched the internet, found the pdf for my system - instructed me on what button to press to pair and get the certificates - it connected to the system and found all the home devices (lights, shades, HVAC temperature control, motion sensors etc.) - it turned on and off my kitchen lights to check that things are working (lol!) I am now vibe coding the home automation master command center, the potential is 🔥.And I'm throwing away the crappy, janky, slow Lutron iOS app I've been using so far. Insanely fun :D :D
claude figured out how to control my oven
795
1,660
437
25,474
3,020,091
Nvidia is buying Groq for two reasons imo.   1) Inference is disaggregating into prefill and decode. SRAM architectures have unique advantages in decode for workloads where performance is primarily a function of memory bandwidth. Rubin CPX, Rubin and the putative “Rubin SRAM” variant derived from Groq should give Nvidia the ability to mix and match chips to create the optimal balance of performance vs. cost for each workload. Rubin CPX is optimized for massive context windows during prefill as a result of super high memory capacity with its relatively low bandwidth GDDR DRAM. Rubin is the workhorse for training and high density, batched inference workloads with its HBM DRAM striking a balance between memory bandwidth and capacity. The Groq-derived "Rubin SRAM" is optimized for ultra-low latency agentic reasoning inference workloads as a result of SRAM’s extremely high memory bandwidth at the cost of lower memory capacity. In the latter case, either CPX or the normal Rubin will likely be used for prefill.   2) It has been clear for a long time that SRAM architectures can hit token per second metrics much higher than GPUs, TPUs or any ASIC that we have yet seen. Extremely low latency per individual user at the expense of throughput per dollar. It was less clear 18 months ago whether end users were willing to pay for this speed (SRAM more expensive per token due to much smaller batch sizes). It is now abundantly clear from Cerebras and Groq’s recent results that users are willing to pay for speed.   Increases my confidence that all ASICs except TPU, AI5 and Trainium will eventually be canceled. Good luck competing with the 3 Rubin variants and multiple associated networking chips. Although it does sound like OpenAI’s ASIC will be surprisingly good (much better than the Meta and Microsoft ASICs).   Let’s see what AMD does. Intel already moving in this direction (they have a prefill optimized SKU and purchased SambaNova, which was the weakest SRAM competitor). Kinda funny that Meta bought Rivos. And Cerebras, where I am biased, is now in a very interesting and highly strategic position as the last (per public knowledge) independent SRAM player that was ahead of Groq on all public benchmarks. Groq’s “many chip” rack architecture, however, was much easier to integrate with Nvidia’s networking stack and perhaps even within a single rack while Cerebras’s WSE almost has to be an independent rack.
169
389
74
2,879
1,152,676
Igor Pener retweeted
As I’m currently testing FSD v14.2.1 using a friend’s car, I wanted to share an early look at my FSD feature progress tracker scorecard. It’s based on a mix of extensive v14 video review and my own hands-on testing. My plan is to wrap up testing and publish a full breakdown later this week highlighting the improvements and what still needs work. The numbers may shift slightly, but I’m confident they’re already close. Quick takeaway: there are major improvements in some of the less-frequent but critical areas like emergency vehicle response, road hazard handling, and robotaxi features such as parking and reversing. However, there’s also a small regression in smoothness, with hesitation in lane changes returning. Let me know if you have any thoughts about certain improvements from the last version or issues still carrying over.
62
59
10
511
90,168
Igor Pener retweeted
Interesting paper: 2-bit LLM in complex numbers. all weights in {±1, ±i} (no zero!) Excited to see this extended to multimodal (audio/video/phase-heavy data is where complex reps should really shine) Paper: arxiv.org/abs/2508.05571 Code & models: github.com/PKULab1806/Fairy-… HF weights: • 700M → huggingface.co/PKU-DS-LAB/Fa… • 1.3B → huggingface.co/PKU-DS-LAB/Fa… #ComplexLLM #iFairy @PKU1898
4
3
23
1,866
Igor Pener retweeted
as much as i hate to admit this
110
437
74
5,303
352,626
$1 of cash on NVIDIA's balance sheet is worth $1. But $1 of revenue at ~50% net margins with a 50x P/E ratio, is worth $25.
Nvidia is investing $100B in its own revenue
37
84
6
1,679
147,738
Chinese scientists have developed the best shortest-path algorithm in 41 years! A team from Tsinghua University has broken Dijkstra’s “sorting barrier” — the first improvement since 1984. The new algorithm runs in O(m \log^{2/3} n) time. Potential applications? Faster shorter waits for route calculations, fewer traffic jams, cheaper deliveries, and more efficient computer networks. And, of course, a need to update computer science curricula :)
188
1,401
135
8,202
1,279,776
Only a barely detectable hint of the "AI fuzz" where different elements blend together and it starts sounding like a low bitrate mp3. This is surprisingly crisp and clear. Elements are well separated. Not as good as a professional mix, but for many use cases... good enough.
Introducing Eleven Music. The highest quality AI music model. - Complete control over genre, style, and structure - Multi-lingual, including English, Spanish, German, Japanese and more - Edit the sound and lyrics of individual sections or the whole song
2
2
12
1,877
Igor Pener retweeted
State space models and RNNs compress history into a constant size state, while attn has KV cache scaling linearly in seqlen. We can instead start from RNNs and let the state size grow logarithmically with seqlen. Feels like a sweet spot. Also beautiful connection to classical algo like Fenwick tree and hierarchical matrices
We know Attention and its linear-time variants, such as linear attention and State Space Models. But what lies in between? Introducing Log-Linear Attention with: - Log-linear time training - Log-time inference (in both time and memory) - Hardware-efficient Triton kernels
11
104
1
938
98,643
Hey kids! Looking to play Fortnite? Try this crypto and stock trading app instead. Rated for ages 4+. Courtesy of Apple App Store ads.
327
287
66
8,759
527,612
🚨🚨🚨 "We found the model attempting to write self-propagating worms, and leaving hidden notes to future instances of itself to undermine its developers' intentions."
How is this not the biggest news story in the world?
917
3,268
1,207
18,005
34,966,206