#real Running, cooking, probability 55%, Bayesian, trading, investing, common sense, logic, data analysis, statistical inference & modeling with common sense
Joined February 2018
- Tweets1.8K
- Following464
- Followers199
- Likes4K
Pinned Tweet
Linear representation is the foundation of most, if not all, models, including architecture of deep learning
Feel free to see it in simple codes I shared for both #julialang and #RStats. You can play around with it too
See images and link: gist.github.com/kar9222/82c7…
Kar Chua retweeted
Alibaba's Qwen3.8 Max scores 56 on the Artificial Analysis Intelligence Index at $1.14 per task, but open weights leader Kimi K3 remains 1 point ahead at 25% lower cost per task ($0.86)
@Alibaba_Qwen has released Qwen3.8 Max, which Alibaba states is a 2.4T total parameter MoE activating 95B parameters per forward pass. Alibaba has announced it plans to release the weights next week, a shift in strategy as it has typically kept its Max class of models proprietary. Once released, Qwen3.8 Max would be ~6x larger than Alibaba's largest open weights release to date (Qwen3.5 397B) and the second largest open weights model behind Kimi K3 (2.8T)
Note: we earlier published results showing Qwen3.8 Max scoring 53 on the Artificial Analysis Intelligence Index. Those runs were affected by intermittent issues on the endpoint we were evaluating, and we have re-run all evaluations on Alibaba's public API endpoint
Key takeaways:
➤ Qwen3.8 Max scores 56 on the Artificial Analysis Intelligence Index, up 10 points from Qwen3.7 Max (46). It is in line with Claude Opus 4.8 (max, 56) and sits second among labs from China, ahead of GLM-5.2 (max, 51) but behind Kimi K3 (max, 57)
➤ Qwen3.8 Max scores 1739 Elo on GDPval-AA, a 468 Elo gain over Qwen3.7 Max. This places it ahead of Kimi K3 (1685), effectively tied with Claude Fable 5 (1743) and GPT-5.6 Sol (max, 1730), and behind only Claude Opus 5 (max, 1852)
➤ Gains over Qwen3.7 Max span agentic evaluations, scientific reasoning and coding: Terminal-Bench v2.1 +6 points, CritPt +7 points, SciCode +4 points and HLE +3 points, with GPQA unchanged. AA-LCR (-2 points) and AA-Omniscience (-10 points, driven by a hallucination rate rising 23% to 40%) regress
➤ Qwen3.8 Max costs $1.14 per Intelligence Index task, more than double Qwen3.7 Max ($0.53), at ~1.3x Kimi K3 (max, $0.86) and ~2x GLM-5.2 (max, $0.57). Cost is driven in part by more turns on agentic evaluations, with GDPval-AA input tokens rising ~15x over Qwen3.7 Max, and output token usage up 45% to 145M
➤ The 𝜏³-Bench Banking result (42%) appears out of distribution. It is a 32 point gain over Qwen3.7 Max and places Qwen3.8 Max ahead of models that outscore it on other evaluations
Key model details:
➤ Size: 2.4T total parameters, ~95B active per forward pass (MoE)
➤ Context window: 1M tokens
➤ Multimodal: text, image and video input with text output
➤ Pricing: $2.00/$6.00 per 1M input/output tokens on the @alibaba_cloud first-party API, with a $0.25 cache hit price. This is lower than Qwen3.7 Max across the board ($2.50/$7.50, with a $0.50 cache hit price)
➤ Availability: Alibaba Cloud first-party API. Alibaba states the weights will be released next week
Kar Chua retweeted
Introducing M2.5, an open-source frontier model designed for real-world productivity.
- SOTA performance at coding (SWE-Bench Verified 80.2%), search (BrowseComp 76.3%), agentic tool-calling (BFCL 76.8%) & office work.
- Optimized for efficient execution, 37% faster at complex tasks.
- At $1 per hour with 100 tps, infinite scaling of long-horizon agents now economically possible
MiniMax Agent: agent.minimax.io
API: platform.minimax.io
CodingPlan: platform.minimax.io/subscrib…
Trying oh-my-opencode /init-deep to generate hierarchical AGENTS.md for main sub-project folders
@opencode @justsisyphus
github.com/code-yeongyu/oh-m…
Using @opencode terminal interface in @Neovim's built- in terminal with opencode.nvim, in Windows Terminal WSL v2. Feels snappy for my small project, and it just works! Feels nice using an open source tool!
Looking to add plugins: opencode.cafe
github.com/NickvanDyke/openc…
Kar Chua retweeted
github.com/code-yeongyu/oh-m…
explore from claude code, oracle from ampcode
basically you can use every model with better tui
why do you pay for others if you can just use it via any providers you wanna use
Kar Chua retweeted
The most INSANE Bohemian Rhapsody flashmob you will ever see!!
With 30 musicians and singers in the streets of Paris 😍
Lead singer: @mickeycallisto
Lead guitar: Olly Pearson
Pop choir: DDKN sisters, @sone_sings and @colinesicre
Counter-tenor solo: @michaelkonesaki
Kar Chua retweeted
In today's article, learn how a chance discussion with the chair of Harvard’s statistics department changed everything about how I think about learning statistics.
Kar Chua retweeted
Upgrade your terminal experience with these CLI tools:
1. fzf
2. ripgrep
3. bat
4. lsd
5. starship
6. just
7. nushell
8. procs
9. k9s
10. difftastic
11. hyperfine
Kar Chua retweeted
Many people assume that LRM reasoning breaks down past a certain "complexity" or "number of steps" threshold. This is incorrect. It breaks down past an unfamiliarity threshold. And that threshold is very low.
There is no limit to the complexity of tasks you can solve with these models, no limit to the number of steps in the reasoning chains they can master -- as long as they have been covered during training / tuning. However, show them something unfamiliar, even very simple and requiring just a handful of reasoning steps (e.g. an ARC 2 task), and they will fail.
The reason you see a steps/complexity threshold when testing on problems like Towers of Hanoi is because these are familiar problems. To see a breakdown you need to turn such a familiar task into a novel one, and one way to achieve this is to scale up the problem variables. The complexity increase is just a roundabout way to generate novelty.
It's not about simple vs complex. It's familiar vs novel. Always has been.
Kar Chua retweeted
Geopolitically, nothing matters more than winning AI.
These tariffs, as constructed, essentially guarantee that America will lose “AI” by making America the most expensive place on earth to build AI datacenters.
The semiconductor exemption was irrelevant for AI. Datacenter semiconductors come into America in finished goods from Taiwan and other Asian countries: servers, storage systems and networking switches. By the time we have developed the capacity to domestically produce these systems, we will have lost the AI race.
Outside of this dynamic, the assumption - or wild gamble - that other large countries would not retaliate and instead come meekly to the negotiating table was erroneous.
That was the off ramp and Bessent did his best to encourage the world to take it. China, understanding how awesome all of this is for them, did not take the off ramp.
Epic win for China.
Best outcome from here - that is in the realm of the possible - would be to quickly cut “deals” with friendly Asian countries - Japan, Taiwan, South Korea, Thailand, Malaysia, Vietnam, Indonesia, etc. Something like a 10% tariff plus true reciprocal tariffs (which would quickly go to zero) would be reasonable.
I say all of this as someone who was open-minded to tariffs that were thoughtfully constructed, gradually phased in and accompanied by massive deregulation. Ideal outcome would have been some combination of reshoring, revenue generation and seeing our largest trading partners lower their barriers to American goods. It is a fact that until yesterday we generally had much lower tariff rates than other countries.
My sense is that is where Bessent was, but he clearly lost an internal battle. Hopefully the market helps the President realize that a course correction is needed and the necessary changes are made relatively quickly.
The sooner, the better.
As a sidenote, I can only imagine how Bessent felt when he understood the way the reciprocal tariffs were calculated, that an LLM had generated the tariffs and that islands inhabited only by Penguins were being tariffed as a result.
Kar Chua retweeted
Use ]] / [[ in :terminal to jump to next/previous shell prompt. ":help terminal-osc133" shows how to create signs, too.
github.com/neovim/neovim/pul…
Note: requires your shell PS1/PROMPT_COMMAND to emit OSC 133 start/end sequences.
Kar Chua retweeted
We strive to ensure that the DuckDB project stays open-source in the long term. That's why we set up the DuckDB Foundation, an Amsterdam-based non-profit organization, in 2021. The Foundation owns the intellectual property of the DuckDB project and enshrines the availability of DuckDB as open-source in its notarized statutes.
In the interest of transparency, we have now published the (English translation of the) Foundation's statutes. Read them at duckdb.org/pdf/deed-of-incor…