Calc Consulting

Congrats to the @cursor_ai team on the launch of Composer 2! We are proud to see Kimi-k2.5 provide the foundation. Seeing our model integrated effectively through Cursor's continued pretraining & high-compute RL training is the open model ecosystem we love to support. Note: Cursor accesses Kimi-k2.5 via @FireworksAI_HQ ' hosted RL and inference platform as part of an authorized commercial partnership.

🚨NEWS: Cursor’s $50B “in-house model” is literally Kimi K2.5 with RL on top. Got caught in 24 hours >be Moonshot AI >spend hundreds of millions training Kimi K2.5 >1 trillion parameters, 15 trillion tokens, agent swarm architecture >beat GPT-5.2 and Opus 4.5 on real benchmarks >open-source it because you believe in the ecosystem >one condition: display “Kimi K2.5” if you make over $20M/month from it >Cursor takes the model >runs RL on coding tasks >ships it March 19 as “Composer 2” >blog post: “continued pretraining + scaled reinforcement learning” >zero mention of Kimi K2.5 >“our in-house models generate more code than almost any other LLMs in the world” >publishes benchmark chart >Composer 2 against Opus 4.6 and GPT-5.4 >uses the chart to justify raising at $50 billion! >less than 24 hours later >kimi dev intercepts the API response >model ID: kimi-k2p5-rl-0317-s515-fast >they didn’t even rename it >Moonshot head of pretraining runs tokenizer test >confirms: identical to Kimi’s tokenizer >publicly tags Cursor’s co-founder: “why aren’t you respecting our license?” >two more Moonshot employees post confirmations >all three posts deleted within hours >legal is now involved >but it gets worse >Cursor had Kimi K2.5 listed as a FREE model in their UI just weeks ago >users were openly using it >Feb 9: “K2.5 was in my model list. I updated and it vanished” >it vanished because Cursor pulled it from the picker, and relaunched it as their own model >Moonshot valuation: $4.3B >Cursor valuation: $50B Absolute state of Cursor.

🚨BREAKING: SUPER MICRO CO-FOUNDER ARRESTED FOR SMUGGLING $2.5B IN NVIDIA GPUs TO CHINA >SMCI co-founder Yih-Shyan "Wally" Liaw arrested today >personally holds $464 MILLION in SMCI stock >charged with smuggling BILLIONS in Nvidia servers to china >used a southeast asian shell company to funnel $2.5B in servers to chinese buyers >$510 million worth shipped in just THREE WEEKS in spring 2025 >built thousands of fake dummy servers to fool U.S compliance auditors >caught on surveillance camera using a HAIR DRYER to swap serial number stickers >coordinated the whole thing over encrypted group chats >SMCI down 12% after hours >faces up to 30 years in federal prison ITS SO OVER…

it was only a matter of time

@naokiss Tomorrow Naked City Talk to Me

Name your top 3 songs from this album (Track List) Is That You? Shandi Talk To Me Naked City What Makes The World Go 'Round Tomorrow Two Sides Of The Coin She's So European Easy As It Seems Torpedo Girl You're All That I Want

I know Silicon Valley startups don't want to hear this..... But the combination of someone in the trades with deep domain expertise and Claude Code will run circles around your generic software. I talked to Cory LaChance this morning, a mechanical engineer in industrial piping construction in Houston. He normally works with chemical plants and refineries, but now he also works with the terminal He reached out in a DM a few days ago and I was so fired up by his story, I asked him if we could record the conversation and share it. He built a full application that industrial contractors are using every day. It reads piping isometric drawings and automatically extracts every weld count, every material spec, every commodity code. Work that took 10 minutes per drawing now takes 60 seconds. It can do 100 drawings in five minutes, saving days of time. His co-workers are all mind blown, and when he talks to them, it's like they are speaking different languages. His fabrication shop uses it daily, and he built the entire thing in 8 weeks. During those 8 weeks he also had to learn everything about Claude Code, the terminal, VS Code, everything. My favorite quote from him was when he said, "I literally did this with zero outside help other than the AI. My favorite tools are screenshots, step by step instructions and asking Claude to explain things like I'm five." Every trades worker with deep expertise and a willingness to sit down with Claude Code for a few weekends is now a potential software founder. I can't wait to meet more people like Cory.

@naokiss A. Aged well.

"Dynasty" Is this album a yes thanks or no thanks? 🅰️ Yes thanks 🅱️ No thanks

@AlexShtf Perhaps we have done something similar I’m not sure Attempted to train models by projecting the weight matrices onto a fixed eigen value structure/distribution we found this works for small problems, but generally speaking is not very good

Well, here it's obviously a fixed point, but not in a sense similar to a regular neural network. An eigenvalue is a solution of a min-max optimization problem, which satisfies the stationarity conditions. So the eigenvector is a fixed point, but the eigenvalue is apparently not.

New post in my "Eigenvalues as models" series. This one asks a practical question: can eigenvalue-based models be made much cheaper to train and evaluate without collapsing into something too simple to be interesting? Dense matrices are expressive but expensive. Fully diagonal ones are cheap but too restrictive. In this post I explore a middle ground that turned out to be much more useful than I expected. It is probably the most implementation-focused entry in the series so far: structured matrices, PyTorch/SciPy plumbing, and experiments. If you care about spectral methods, differentiable numerical linear algebra, or unusual tabular model classes, this is post and the entire series is for you: alexshtf.github.io/2026/03/15/Spe…

@naokiss God of Thunder -- gotta mention: Goin' Blind Got to Choose incredible range

Name just one KISS song that title begins with the letter "G"?

@AlexShtf Note that the trace log symmetry admits a trivial solution which corresponds to a gaussian fixed point. In that solution, the gap between all the eigenvalues is the same since all the eigenvalues are 1.0

That's interesting. So here instead of layer weights I have matrices in the linear pencil. These are also model parameters that are matrices. But I think in my case whats important is the eigenvalue gap. I want eigenvalue gaps to be as large as possible. This is not easy to achieve through theory.

We have observed that the eigenvalue spectrum of layer weight matrices tend to be heavy, tailed power law, and follow the trace log condition on the eigenvalues in the power law tail The student teacher construct is theoretical machinery You don’t actually need to have a teacher and a student that works or just any model of your training. The alternative way of looking at this is that you expect during training the system to approach a fixed point therefore it is amenable to analysis using re-normalization group (RG) under RG the trace log condition is the underlying dynamical symmetry or Ward identity. using this identity It’s fairly easy to prove the distribution is heavy tailed and satisfies the symmetry that it will converse at the fixed point to a power law with PL exponent alpha=2

@CalcCon WDYM? I don't have teacher and student. So correlation between what and what? Please explain.

@TrungTPhan I still remember the day I saw that in the local theatre. There were maybe 10 people there.

James Cameron has great explainer on how he made "The Terminator" (1984) for cheap: ▫️approach the story as a low-budget horror slasher film  ▫️find areas with brightest street lights, to save money on lighting ("used car lots were great because they cast enough light from their floodlights onto the streets") ▫️the casting of Arnold (he hadn't really broken through yet and was only paid $75,000 which worked out to $1,293 per word) The horror approach also worked because they could shoot a lot of the film guerilla style (no permits) in dark alleyways and parking garages. Smartly, Cameron layered on the sci-fi backstory to make it higher concept (leading to the glorious T2 sequel). There were also a lot of indoor shots (apartment, store, bar) that didn’t need permits. At the time, 30-year old Cameron was living so frugally that his mom was sending him "2-for-1 Big Mac coupons" from McDonald's because she didn't think he was eating enough.  The film ultimately made $80 million globally on a $4.3 million budget, launching Cameron and Arnold's Hollywood careers.

@nickshirleyy @nearlydaniel We haven't forgotten

@nearlydaniel Imagine being the governor (or his staff) and thinking this is the way to lead by example. Vile and weak people

Unbelievably radicalizing exchange The fact that Newsom (correctly?) thinks this will gain him points is horrific Any competent politician (forgive the oxymoron) would tighten budgets at every turn To openly mock those exposing the disgusting fraud and abuse ... Insane.

@ksorbs Her story needs to be told.

Canada is gone.

WeightWatcher XGBoost Update 1600 rounds of XGBoost on some random model...I extract the correlations into a matrix, apply weightwatcher..slowly it converges to the perfect alpha=2 Exactly as predicted by theory

@naokiss Superstar

Describe Paul Stanley in ONE word