Samir Kumar retweeted
Rumors I’ve been hearing. I’ll start with OpenAI, then Anthropic.
OpenAI: The internal model’s reasoning jump comes from more aggressive RL, including training in games and simulated environments. Strategies learned there transfer to research mathematics and other fields. Navier–Stokes and ETH show the scale of the improvement. The clue is in the Synthetic RL hiring: self-play, simulators, and synthetic feedback.
Anthropic: Same direction with its internal Claude model. It’s trained on long, difficult tasks in demanding environments, and the reasoning skills transfer to other fields. Its ETH result shows how far the improvement has gone. The earlier Universes listing and the RL hiring around mathematical reasoning align with that.
Musk’s Digital Optimus hiring explicitly calls for measuring transfer from games into general computer use. That’s why his game-playing AI push matters: games develop abilities that work far beyond the games themselves.
The interesting part is where those abilities show up. Each lab’s internal model applies reasoning developed through interactive training to mathematical problems that had resisted decades of work.
OpenAI: The internal model’s reasoning jump comes from more aggressive RL, including training in games and simulated environments. Strategies learned there transfer to research mathematics and other fields. Navier–Stokes and ETH show the scale of the improvement. The clue is in the Synthetic RL hiring: self-play, simulators, and synthetic feedback.
Anthropic: Same direction with its internal Claude model. It’s trained on long, difficult tasks in demanding environments, and the reasoning skills transfer to other fields. Its ETH result shows how far the improvement has gone. The earlier Universes listing and the RL hiring around mathematical reasoning align with that.
Musk’s Digital Optimus hiring explicitly calls for measuring transfer from games into general computer use. That’s why his game-playing AI push matters: games develop abilities that work far beyond the games themselves.
The interesting part is where those abilities show up. Each lab’s internal model applies reasoning developed through interactive training to mathematical problems that had resisted decades of work.
138
222
4,369
562,387
Samir Kumar retweeted
I think our industry partners rather doubted we would get anywhere near these numbers but PsiQuantum team made this look easy. This chip enables all-optical switching of network traffic without pluggable transceivers, moving mirrors, or amplifiers to compensate for loss.
Google are still the only company brave enough to do OCS at any meaningful scale — huge credit to them for slamming free-space optics (!) in a datacenter rack. This photonic chip achieves similar power, throughput, and architectural advantages without moving parts, and in a much smaller package.
64-port strictly non-blocking silicon will go into test this year, 72-port device tapes out soon. All of these chips will run well under 3dB loss thanks to ludicrously good passives (<1mdB splitters and crossings, 1dB/m waveguides) and advanced edge coupling (~5X better than industry). No need to put an amplifier on every fiber.
All of this is done using simple millisecond-rate thermal phase shifters, but lurking in the background is the prospect of fast switching with barium titanate — PsiQuantum already builds BTO switches that switch up to 64 ports in a nanosecond, and — somewhat in contrast with TFLN – we can fit a ton of switches on a reticle. This goes far beyond current industry demand but will unlock categorically new architectures for training and inference.
We’re excited to make our first serious foray into classical computing hardware and believe this product will make the architectural benefits of optical circuit switching simpler and more accessible to the industry.
We also would not be doing this if it were not extremely well-aligned with PsiQuantum’s mission. The OCS chip is strikingly similar to parts that we already make for pump laser distribution in the photonic quantum computer — so similar in fact that it only took a dozen engineers a couple weeks to design the part. We did not modify the photonics process, PDK, or packaging flow to achieve these results.
Now, through the process of building hundreds of thousands of these chips for industry, we’re also going to improve the yield, reliability, and supply chain for parts that we need for utility-scale photonic quantum computing.
This is a direct realization of the founding thesis of PsiQuantum — that by aligning ourselves with the prevailing wind of the conventional computing industry, fabs, OSATs, CMs etc, we can de-risk and accelerate the hard journey to a useful fault-tolerant quantum computer.
Google are still the only company brave enough to do OCS at any meaningful scale — huge credit to them for slamming free-space optics (!) in a datacenter rack. This photonic chip achieves similar power, throughput, and architectural advantages without moving parts, and in a much smaller package.
64-port strictly non-blocking silicon will go into test this year, 72-port device tapes out soon. All of these chips will run well under 3dB loss thanks to ludicrously good passives (<1mdB splitters and crossings, 1dB/m waveguides) and advanced edge coupling (~5X better than industry). No need to put an amplifier on every fiber.
All of this is done using simple millisecond-rate thermal phase shifters, but lurking in the background is the prospect of fast switching with barium titanate — PsiQuantum already builds BTO switches that switch up to 64 ports in a nanosecond, and — somewhat in contrast with TFLN – we can fit a ton of switches on a reticle. This goes far beyond current industry demand but will unlock categorically new architectures for training and inference.
We’re excited to make our first serious foray into classical computing hardware and believe this product will make the architectural benefits of optical circuit switching simpler and more accessible to the industry.
We also would not be doing this if it were not extremely well-aligned with PsiQuantum’s mission. The OCS chip is strikingly similar to parts that we already make for pump laser distribution in the photonic quantum computer — so similar in fact that it only took a dozen engineers a couple weeks to design the part. We did not modify the photonics process, PDK, or packaging flow to achieve these results.
Now, through the process of building hundreds of thousands of these chips for industry, we’re also going to improve the yield, reliability, and supply chain for parts that we need for utility-scale photonic quantum computing.
This is a direct realization of the founding thesis of PsiQuantum — that by aligning ourselves with the prevailing wind of the conventional computing industry, fabs, OSATs, CMs etc, we can de-risk and accelerate the hard journey to a useful fault-tolerant quantum computer.
AI is pushing computing infrastructure to its limits. Our photonics can help.
Today, PsiQuantum is unveiling its industry-leading Optical Circuit Switch initiative with an 8×8 OCS demo built on the company’s low-loss silicon photonics platform—the same platform we’ve spent ing for quantum computing.
11
22
155
10,165
Samir Kumar retweeted
AI is pushing computing infrastructure to its limits. Our photonics can help.
Today, PsiQuantum is unveiling its industry-leading Optical Circuit Switch initiative with an 8×8 OCS demo built on the company’s low-loss silicon photonics platform—the same platform we’ve spent ing for quantum computing.
Today, PsiQuantum is unveiling its industry-leading Optical Circuit Switch initiative with an 8×8 OCS demo built on the company’s low-loss silicon photonics platform—the same platform we’ve spent ing for quantum computing.
21
46
387
48,806
Samir Kumar retweeted
Today, we’re introducing the next chapter of Concurrence and the world’s greatest leap toward Superintelligent Care.
When I was 13 I lost my mom to cancer that should’ve been caught earlier.
We’re building a system that sees every patient’s future in time to change the outcome
When I was 13 I lost my mom to cancer that should’ve been caught earlier.
We’re building a system that sees every patient’s future in time to change the outcome
164
147
1,476
1,532,068
Samir Kumar retweeted
Introducing Echo-3. A state-of-the-art world model.
Echo-3 turns phone captures into photorealistic, metric-scale 3D worlds. From a single room to multi-floor buildings, outdoor and public spaces.
Available today.
Echo-3 turns phone captures into photorealistic, metric-scale 3D worlds. From a single room to multi-floor buildings, outdoor and public spaces.
Available today.
98
238
1,973
306,201
Samir Kumar retweeted
According to the book, Kushner got Tillerson fired for a $50 million UAE payment because Tillerson supported Qatar. This panicked Qatar because they realized Kushner was a pay-to-play operative, so they bought Kushner's New York City apartments, which were in the red, hoping this would help.
This book is really damning for Kushner. It claims the UAE paid Kushner $50 million to get Rex Tillerson fired as Secretary of State.
161
4,567
13,249
548,221
Samir Kumar retweeted
We're excited to introduce AW-0: Agentic World Action Models! 🤖 (0/8)
The first World Action Model that plans in code, predicts the future, and controls the robot—all in one generative model.
AW-0 jointly generates:
🧠 Code - explicit skill calls, object arguments, and video - predictive visual dynamics
🦾 Robot actions - continuous motor control
All three are generated together through joint denoising, letting semantic plans directly shape both predicted futures and robot behavior, with:
🤝 Hybrid execution — generated programs can also run directly through perception and control APIs
⚡ Just 40 ms of additional inference latency over the base WAM on an RTX 5090
📚 Pretrained on ~3,000 hours of robot + human demonstrations
One model for planning, prediction, and execution.
🌐 Project: agentic-wam.github.io
📄 Paper: arxiv.org/abs/2610.30136…
💻 Code: github.com/ZhuoyangLiu200…
🤗 Model: huggingface.co/miniFranka/age…
🧵 How it works—and what we found ↓
The first World Action Model that plans in code, predicts the future, and controls the robot—all in one generative model.
AW-0 jointly generates:
🧠 Code - explicit skill calls, object arguments, and video - predictive visual dynamics
🦾 Robot actions - continuous motor control
All three are generated together through joint denoising, letting semantic plans directly shape both predicted futures and robot behavior, with:
🤝 Hybrid execution — generated programs can also run directly through perception and control APIs
⚡ Just 40 ms of additional inference latency over the base WAM on an RTX 5090
📚 Pretrained on ~3,000 hours of robot + human demonstrations
One model for planning, prediction, and execution.
🌐 Project: agentic-wam.github.io
📄 Paper: arxiv.org/abs/2610.30136…
💻 Code: github.com/ZhuoyangLiu200…
🤗 Model: huggingface.co/miniFranka/age…
🧵 How it works—and what we found ↓
25
49
316
24,854
Samir Kumar retweeted
🔴 Today , they released the first edition of the book "Princes and Thieves: Jared Kushner's and Mohammed bin Zayed's Secret Conspiracy to Reshape the Middle East." These are the most important points from its first chapter, where investigative journalist Matthew Cole reveals ting that brought together Sheikh Mohammed bin Zayed and Jared Kushner in New York in December 2016, just before Trump's inauguration as president.
This chapter outlines how the UAE leadership bypassed the Obama administration to reach directly into the inner circle of the president-elect, leveraging its network of financial and investment ties. Most crucially, the book uncovers what was discussed during the meeting regarding Qatar, Iran, the Muslim Brotherhood, and Israel, along with the priorities that Abu Dhabi wanted the new administration to adopt. It also traces the role Kushner played in opening unconventional channels of communication, and how that meeting, according to the author's account, laid the foundation for a political relationship whose effects extended into subsequent years.
🔷 The book reveals a meeting that brought together Jared Kushner and Mohammed bin Zayed in the presidential suite at the Four Seasons Hotel in Manhattan, after Trump's election victory and before his inauguration. Bin Zayed arrived with a delegation that included dozens of people, among them armed individuals. The meeting came after he decided to cancel a farewell banquet that had been scheduled with President Barack Obama at the White House.
🔷 The Emiratis offered no explanation for canceling his meeting with Obama, and the U.S. administration only learned of his New York visit after the State Department received notifications of his diplomatic travel itinerary. The author describes this action as a clear circumvention of diplomatic norms and a signal of the UAE leadership's shift in focus from the outgoing administration to the personal circle of the president-elect.
🔷 Mohammed bin Zayed enlisted the help of a New York-based hedge fund manager with financial ties to the UAE and a personal friendship with Kushner to arrange the meeting. This serves as an indication of how the UAE uses its investment network to access influential American political figures.
🔷 Kushner informed bin Zayed that the Middle East would be the top priority of the Trump administration's foreign policy, and that he would personally take charge of the diplomatic file on the Israeli-Palestinian conflict. Trump had selected him to handle these matters despite his lack of traditional diplomatic experience. Meanwhile, bin Zayed laid out his grievances to Kushner about U.S. policy over the previous eight years, viewing Obama's support for the Arab Spring uprisings—particularly the protests that toppled Egyptian President Hosni Mubarak—as a betrayal of a key U.S. strategic ally. He explained that Washington's abandonment of Mubarak had heightened Gulf regimes' concerns about the reliability of American guarantees.
🔷 He further argued that the Obama administration's decision to end the U.S. war in Iraq had created a strategic vacuum that enabled the Islamic State to expand and control vast territories in Iraq and Syria. He presented this outcome as evidence of the weakness in U.S. policy in the region. He also told Kushner that the UAE faces two enemies threatening its existence: the Muslim Brotherhood and Iran. He viewed the first threat as political Islam, while the second stemmed from Iran's regional influence and its network of armed groups linked to Tehran.
🔷 He accused Qatar of providing significant funding to the Muslim Brotherhood and using Al Jazeera to stoke regional tensions and support extremist Islamist currents, according to his narrative. He informed Kushner that he hoped the new U.S. administration would address the Qatari problem. This passage reveals that the Qatar file was already on the table in UAE communications with the Trump in Dec. 2016.
This chapter outlines how the UAE leadership bypassed the Obama administration to reach directly into the inner circle of the president-elect, leveraging its network of financial and investment ties. Most crucially, the book uncovers what was discussed during the meeting regarding Qatar, Iran, the Muslim Brotherhood, and Israel, along with the priorities that Abu Dhabi wanted the new administration to adopt. It also traces the role Kushner played in opening unconventional channels of communication, and how that meeting, according to the author's account, laid the foundation for a political relationship whose effects extended into subsequent years.
🔷 The book reveals a meeting that brought together Jared Kushner and Mohammed bin Zayed in the presidential suite at the Four Seasons Hotel in Manhattan, after Trump's election victory and before his inauguration. Bin Zayed arrived with a delegation that included dozens of people, among them armed individuals. The meeting came after he decided to cancel a farewell banquet that had been scheduled with President Barack Obama at the White House.
🔷 The Emiratis offered no explanation for canceling his meeting with Obama, and the U.S. administration only learned of his New York visit after the State Department received notifications of his diplomatic travel itinerary. The author describes this action as a clear circumvention of diplomatic norms and a signal of the UAE leadership's shift in focus from the outgoing administration to the personal circle of the president-elect.
🔷 Mohammed bin Zayed enlisted the help of a New York-based hedge fund manager with financial ties to the UAE and a personal friendship with Kushner to arrange the meeting. This serves as an indication of how the UAE uses its investment network to access influential American political figures.
🔷 Kushner informed bin Zayed that the Middle East would be the top priority of the Trump administration's foreign policy, and that he would personally take charge of the diplomatic file on the Israeli-Palestinian conflict. Trump had selected him to handle these matters despite his lack of traditional diplomatic experience. Meanwhile, bin Zayed laid out his grievances to Kushner about U.S. policy over the previous eight years, viewing Obama's support for the Arab Spring uprisings—particularly the protests that toppled Egyptian President Hosni Mubarak—as a betrayal of a key U.S. strategic ally. He explained that Washington's abandonment of Mubarak had heightened Gulf regimes' concerns about the reliability of American guarantees.
🔷 He further argued that the Obama administration's decision to end the U.S. war in Iraq had created a strategic vacuum that enabled the Islamic State to expand and control vast territories in Iraq and Syria. He presented this outcome as evidence of the weakness in U.S. policy in the region. He also told Kushner that the UAE faces two enemies threatening its existence: the Muslim Brotherhood and Iran. He viewed the first threat as political Islam, while the second stemmed from Iran's regional influence and its network of armed groups linked to Tehran.
🔷 He accused Qatar of providing significant funding to the Muslim Brotherhood and using Al Jazeera to stoke regional tensions and support extremist Islamist currents, according to his narrative. He informed Kushner that he hoped the new U.S. administration would address the Qatari problem. This passage reveals that the Qatar file was already on the table in UAE communications with the Trump in Dec. 2016.
30
916
1,734
64,594
Samir Kumar retweeted
Meet the world-class research team behind SKY-001, the first Engineering World Model.
Solving complex design and engineering problems in the physical world is orders of magnitudes more difficult than LLM agents using next token prediction. Our team at @skyfallai decided to take veryone else out there in order to achieve the research breakthrough.
Last week, we shared how SKY-001 stacks up against frontier LLMs, showing the clear advantage World Models have on complex real-world tasks like engineering design.
Here is a sneak peek video where you can learn directly from the researchers about what makes our approach unique, what training and data collection paradigms were used, and why we believe World Models are the next frontier in AI.
I’m incredibly proud to work alongside this team.
Solving complex design and engineering problems in the physical world is orders of magnitudes more difficult than LLM agents using next token prediction. Our team at @skyfallai decided to take veryone else out there in order to achieve the research breakthrough.
Last week, we shared how SKY-001 stacks up against frontier LLMs, showing the clear advantage World Models have on complex real-world tasks like engineering design.
Here is a sneak peek video where you can learn directly from the researchers about what makes our approach unique, what training and data collection paradigms were used, and why we believe World Models are the next frontier in AI.
I’m incredibly proud to work alongside this team.
29
26
229
23,390
Samir Kumar retweeted
This post talks about companies getting to $1M/month which is easy but tenuous. How do you get to $100M/month as quickly as possible so you can thrive? A lot of inference infrastructure needs to be built because the demand is there, but financiers are skittish to underwrite it. The thesis of heterogeneous multi-silicon inference we have at Parasail is to dynamically match requests to the best piece of hardware (both request routing and disaggregated prefill/decode) to get 4-5X more efficiency. This is important because hardware financiers are not clear on how long term dynamics play out with nvidia inference, but they will underwrite a massive improvement in $/MW because you’re no longer in commoditized Nvidia land, payback periods are much faster, and cost of capital impact is blunted.
There is an absolute barrage of inference providers popping up that get to $1M/mo extremely quickly.
This is the rough economics of inference providers:
Even if you rent on-demand B300s at $8/GPU-hr, many open source models (GLM 5.3, Kimi K3, DeepSeek 4.1 Flash / Pro) can get easonable throughputs (100tok/s). DeepInfer from DeepSeek directly shows they can get $38/GPU-hr on off-peak hours. Every chip / model has a different throughput / interactivity curve, but there is a lot of arbitrage between even high rental spot prices and token prices. New chips (Vera Rubins) also change this meaningfully. Your ability to be making gross margin relies on your utilization and how good your stack is.
Gateways like OpenRouter, Vercel and more help solve that problem by aggregating demand to you. May inference providers discount their tokens to get longer term commits too. I've even heard stories of selling back Chinese model tokens to Chinese labs!
One thorn in the economics of these players is that many Chinese open source model labs are now asking for rev share if you're at a certain scale, espp for pre-release access to their models for day 1 support. That further pushes the threshold to be in the money.
The barrier to entry seems low and falling. There are even stories of smaller ones who can "vibe code" a way to stand up OSS models on a cluster they rent. And as a result, we're seeing 10s of such companies pop up almost every week and immediately skyrocketing revenue at low-ish gross margin and raising rounds on the back of that.
Unsurprisingly, because the actual service provided seems commodity, this eventually drives up compute prices and drives down token prices. In theory, it would seem like neoclouds who have the lowest cost of compute and themselves have started inference businesses would win the race to the bottom (Crusoe Coreweave Nebius and GMI all have inference businesses), with larger players (Modal Fireworks Baseten) being second because they are well capitalized to make bigger compute commits cheaply and have demand (high utilization), but the reality is the pie is growing so quickly that even upstarts are absolutely crushing it and growing aggressively (inference dot net, DeepInfra, Wafer, Relace, Morph).
Strategies vary across the board: neoclouds want to serve inference, inference players want to be neoclouds and do higher margin reserved instances / post training, and some want to go heterogeneous hardware or creatively flexible compute constructs.
The risks are: do open source models continue to be competitive? Will their creators rent seek more? Will their shape change compute needs? How do different chips change economics? Who gets compute in a supply crunch?
Below is a aggressive sketch of how gross margin scales with utilization for various models / chip sets.
For now, there are many inference providers and they’re all ripping!
1
1
7
546
Samir Kumar retweeted
So they finally caught up to Jared. Let’s hope they do something about it.
146
2,441
5,423
91,699
Samir Kumar retweeted
Whoever works on OneDrive search at Microsoft should be fired and possibly tortured on their way out
66
36
1,313
40,528
Samir Kumar retweeted
So the first thing I'm crazy excited about is the boy genius Elliot Arledge @elliotarledge, the creator of this CUDA course and really a self-made person: someone who eschewed a college-classic education for teaching, as a renegade, state-of-the-art CUDA engineering to millions of people via Free Code Camp. He does kind of 22-hour coding sessions and marathons in a room heated by melting GPUs and is heroically devoting himself to unlocking absolutely every chip's full capacity in the great engineering challenge of Ignition at @infinity_ai_ .
There's something to be said for people who continue to pursue greatness in an era where greatness is taboo. Elliot stands proud as a person who I expect will achieve tremendous great things at Infinity and beyond.
There's something to be said for people who continue to pursue greatness in an era where greatness is taboo. Elliot stands proud as a person who I expect will achieve tremendous great things at Infinity and beyond.
timelapse #83 (22 hrs):
- it was very easy to dive super deep into anything i needed to (this is what i focused on today because not all days are like this)
- finding the grok code fast 1 + grok 4 for deep thinking and verification combo to be super useful in cursor. speed was myself spending many more mental clock cycles in a 24 hr period
- had to pull out qwen3-next’s gated deltanet + linear attention from bleeding edge hf transformers to begin implementing a multi-gpu fp8 trainer from scratch. this is so damn bleeding edge and i underestimated how much effort this has and will require lol
- lots of diet coke and oats
- shipped the template which the core chapters of my book will be built on: github.com/infatoshi/naiv…
- all im missing now is flash attention 1/2 mastery (fa2 tmrw), intuition on making topk faster (for arbitrary row length), what i should and shouldnt teach in cutlass/cute, hopper/blackwell gemm kernel mastery (down to fp4) — shoutout to @pranjalssh for making this easier for me. his blog post is amazing
- caught up w/ @matiroy
- im feeling great mentally but not so good physically as im writing this and about to pass out
5
10
176
15,168
Samir Kumar retweeted
Proud to have the highest tps on @OpenRouter for @MiniMax_AI M3.1 Flash
0
4
6
262
Samir Kumar retweeted
Among the gems that the Dutch government found on Jared Kushner's phone, payments to a known hitman, a warlord, hundreds of millions from bank accounts in Central Asia..on and on..
Jared Kushner’s Hacked Phone Revealed His Secret Negotiations—Including More Than $700 Million In Transfers vanityfair.com/story/jared-ku…
149
5,802
22,163
372,959
Samir Kumar retweeted
The GRU hacked into Jared Kushner's iPhone. Dutch military intelligence (MIVD) then recruited a GRU spy who gave them access to the hacked phone. What MIVD discovered was that Kushner had access to two bank accounts --one Azerbaijani, the other Israeli --where Saudi Crown Prince Mohammed bin Salman and Emirati ruler Mohammed bin Zayed deposited money. It sounds like a slightly far-fetched episode of "Slow Horses," but it's real. And it proves just how eye-poppingly corrupt this administration was and continues to be as its "special envoy" decides the fate of tens of millions in Europe in the Middle East. Simply incredible reporting by Matthew Cole, drawn from his new book on Kushner. vanityfair.com/story/jared-ku…
1,456
24,382
77,713
6,539,067
Samir Kumar retweeted
BREAKING: Jared Kushner’s hacked iPhone revealed that Kushner promised MBS and MBZ he would push their agenda with Trump if they participated in a secret fund, with the two royals initially discussing over $10 billion before ultimately placing at least $2 billion into accounts for him, according to a Vanity Fair excerpt from Princes and Thieves.
The revelation lands as Kushner remains deeply involved in Iran war diplomacy, while Tehran has separately accused him and Steve Witkoff of pocketing $9 billion through market manipulation.
The revelation lands as Kushner remains deeply involved in Iran war diplomacy, while Tehran has separately accused him and Steve Witkoff of pocketing $9 billion through market manipulation.
324
6,395
28,019
2,312,212
Samir Kumar retweeted
We launched the first Engineering World Model, SKY-001, this week and it already outperforms all the frontier LLMs in CAD and BIM in speed, cost and precision.
Here’s an example of how our model compares to OpenAI’s Astra, Anthropic’s Fable 5.1 and Google’s Gemini 4.
These by accident. Read our technical blog to find out how we built SKY-001 (🔗 below).
Here’s an example of how our model compares to OpenAI’s Astra, Anthropic’s Fable 5.1 and Google’s Gemini 4.
These by accident. Read our technical blog to find out how we built SKY-001 (🔗 below).
48
81
1,051
63,567
Samir Kumar retweeted
Effective Immediately, all organizations using PHP for their systems are advised to be prepared for an update addressing two critical vulnerabilities. PHP is widely used for web applications. These vulnerabilities often provide a direct path to the server's file system or database. If an immediate update is not feasible, it is recommend that organizations take the system offline until it can be fully mitigated. NSA will follow up shortly with more specific guidance.
1,296
10,837
70,321
12,590,551