Cheng (@chumsdock) header
Cheng (@chumsdock) profile picture

Design with rationality. https://xitter.gq/t.co/UBMlLQnWoe

Beijing
Joined January 2007
Photo by Cheng (@chumsdock)Photo by Cheng (@chumsdock)Photo by Cheng (@chumsdock)Photo by Cheng (@chumsdock)Photo by Cheng (@chumsdock)Photo by Cheng (@chumsdock)Photo by Cheng (@chumsdock)Photo by Cheng (@chumsdock)Photo by Cheng (@chumsdock)Photo by Cheng (@chumsdock)
Cheng retweeted
using carbon with 4 hydrogen subagents. i call it... methane.
44
73
18
1,499
28,973
Cheng retweeted
we got LLMs to do weight space programming on a CNN for CIFAR-10, and found lots of cool behaviors in the reasoning traces! models invented their own training data, reasoned about images through ASCII art, and learned to map features to labels entirely in-context. had fun experimenting with Ripple-65B and other LLMs alongside @mridul__agarwal @Vedang_A at the hackathon, we wrote up what we found 🧵
Last Friday we hosted the Ricursive Intelligence After Dark Hackathon, our first hackathon centered on our open-weight model release, Ripple-65B Preview. We brought together an incredible group of researchers, engineers, and builders to push the model in new directions. We saw some amazing projects spanning everything from CUA and Weight-Space Programming to model watermarking. Huge thanks to everyone who came out, built something, and made the night memorable! This is just the beginning!
3
3
2
43
2,298
Cheng retweeted
For a "software factory" to scale, you need a set of strong primitives. I feel like this auto-approving feature is an interesting primitive worth learning about.
40
23
2
229
23,161
Cheng retweeted
Great paper from Google on making ML experiments reproducible for coding agents. (bookmark it) If you run agents on research code, this one is worth your time. They find that refined artifacts raise agent reproduction success from 76.4% to 92.6% across four models and cut API spending by 19.7%. ReproBridge turns one successful reproduction into a package the next agent can reuse. The package holds the cleaned code, the environment setup, scripts that launch each experiment, and a written guide. A new agent reruns the experiment from that package. A separate auditor then checks that the runs actually happened. The benchmark covers 1,029 tasks from 36 papers, each with five human-verified reference scores. The gap between the best and worst model drops from 21.6 to 3.9 points. Removing the code or the environment setup causes the largest losses. If your agents keep failing to rerun a repo, save the setup steps from the first successful run before trying a bigger model. arxiv.org/abs/2611.08452 Chat with Paper: academy.dair.ai/papers/repro…
83
24
10
166
13,577
everyone is confused about google model naming, but it's really quite simple. Bard > Gato > LaMDA > LaMDA 2 > PaLM > PaLM 2 > Gemini 1.0 Ultra > Gemini 1.0 Pro > Gemini 1.0 Nano > Gemini 1.5 Pro > Gemini 1.5 Flash > Gemini 2.0 Flash > Gemini 2.0 Flash Thinking > Gemini 2.5 Pro > Gemini 2.5 Flash > Gemini 2.5 Flash-Lite > Gemini 3 Pro > Gemini 3 Flash > Gemini 3.1 Pro > Gemini 3.1 Flash-Lite > Gemini 3.5 Flash > Gemini 3.5 Flash-Lite > Gemini 3.6 Flash > Gemini 3.7 Flash > Gemini 3.8 Flash > Gemini 3.8 Flash-Lite > Gemma 1 2B > Gemma 1 7B > Gemma 2 2B > Gemma 2 9B > Gemma 2 27B > Gemma 3 270M > Gemma 3 1B > Gemma 3 4B > Gemma 3 12B > Gemma 3 27B > Gemma 3n E2B > Gemma 3n E4B > Gemma 4 E2B > Gemma 4 E4B > Gemma 4 12B > Gemma 4 26B A4B > Gemma 4 31B > Gemini 4 Argon > Gemini 4 Carbon > Gemini 4 Hydrogen > Gemini 4 Neon
58
88
20
1,509
70,558
I think the parsimonious explanation of all evidence to date is this: - Alice Huang is not a spy - the MSS was NOT involved in MegaSpeed or in smuggling compute - major Chinese companies generally prefer to legally rent compute or build overseas datacenters, and need external prodding to do riskier stuff - the policy of the PRC is AGAINST smuggling compute, refusing H20s and then H200s was not some "bluff" of a paper tiger, but the straightforward expression of their cost-benefit calculation with regard to EV of domestic development velocity (which in my opinion is incorrect). In fact, if you think they're rationally acting to maximize compute stockpile, they have zero reason to not accept H200s after the licenses were granted, they can walk and chew bubble gum at the same time, import legal and smuggle illegal chips in parallel. They did neither.* - JENSEN HUANG THE CEO OF NVIDIA, personally, has been very interested in selling GPUs to China, for years, by hook or by crook, by modifying chip designs to pass right below the permissible cutoff, by forging ties with Chinese and Southeast Asian businesspeople including Alice whom he's known for a while, by lobbyism and psyops, and by leveraging his connections in the admin. It was, indeed, all Jensen, which is why Jensen had to go. My presumption is that it was not so much to increase shareholder value or Jensen's own wealth, and certainly not out of some racial loyalty (as I've argued repeatedly, Jensen, being Taiwanese, is more likely interested in defusing the Taiwan scenario than in propping up the Mainland), but to maintain a stake for Nvidia in the Chinese market and hamper the development of by-default-noncompetitive Chinese AI hardware. This makes sense if you consider his priors, which may seem outlandish in the AGI-pilled or natsec crowd but are mundane in his milieu. Huang is very invested (in all ways, but also psychologically and even genetically, having given positions to his children) in the lasting global success of Nvidia. He believes AI to be a "normal technology", not a supergiganuke that can grant or deny eternal hegemony to the US (this belief may be argued to be downstream of other motivations). He is a consummate capitalist who thinks that eventually the political climate may change and US and Chinese companies will compete as compute providers on the open global market, and he does not intend to cede any segment of that market to Huawei et al., so his best play was strangling them in their cribs, making sure every top-tier Chinese AI company remains dependent on Nvidia and the only feedback Huawei receives is from fellow entity list members like iFlyTek. Whatever you think of all these opinions, they are perfectly aligned both with Jensen's public stances and his revealed private actions. *Generally, Americans are always wrong about this kind of shit, they incorrectly imagine China to be a cohesive hivemind with Beijing as the brain bug, and not a very smart brain bug but one that's desperately terrified of them and will do some childishly transparent Sun Tzu plays to deceive them, such as not accepting legal imports to "distract" the investigation or "send a signal" or whatever (would not work even on an individual patriotic commenter on X). It's very unseemly and disappointing. I conclude it's some antisocial psychological defect/pure ideology on the part of Americans, maybe downstream of the damage they receive when brainfucked into the esoteric neocon club, and they are not capable of debating it in good faith. All these burger "eggsperts" like @hamandcheese or @niubi or @jordanschneider will insist I'm wrong or another CCP agent, but their stories fall apart if you look at the numbers (they'll try to get you to focus on some bullshit like 100K chips over 3 years here and 30K chips deployed in Malaysia there, and lie by omission about the success of the childish smuggling-in-the-shadows scheme that I've already covered). That's my thinking about this, as of 18th Nov 2026.
29
76
11
766
28,926
Astra 6.2 和 Gemini 4 Carbon 都挺好的,Opus 5.5 虽然也擅长干活,但是很多问题改来改去没改好的时候,可能让你PUA自己就是这样的,其实换个模型就解决了。
242
DeepMind should replace coding interviews with asking candidates to go through the periodic table from memory
6
1
70
2,668
Ownfeed 马上要上一个超牛的新功能!! X 的 For you 越来越乱了,内容实在是垃圾。 现在可以用 Ownfeed 在 X 里加一栏自己的 feed:说一句你想看什么,它会按对你的了解,帮你挑出真正想看的高质量内容。 比如我自己想看 AI,就加了一栏「AI Fomo」,有了一个只属于我的 AI 信息流。 你们想看什么?欢迎都来试试!!(什么都可以哦🐶
7
2
1
30
35,029
This may be the best @threejs website ever. The quality is insane.
4
4
2
68
4,594
Cheng retweeted
idk why people have so much trouble wrapping their head around agent permissions. An agent is just another principal. We’ve already invented RBAC, scopes, audit logs, all those things apply to agents too
19
7
4
133
9,562
Clever, clever cover from @TheEconomist. AI as a skyscraper, propped up by money, someone familiar tumbling out of the bottom. Crash or just the next turn? They don't say. No one really knows. In medicine we're only just starting to build, but it's moving fast. Interesting times ahead for all of us.
1
3
6
1,442
Cheng retweeted
Anthropic's claim that Claude ported its inference stack to AMD MI355X GPUs in just 3 days with zero human code edits sounds impressive, but there's some important fine print. Two details stand out: 1. The work was done by an internal Claude Mythos Preview, not a publicly available model. 2. It only covered prompt processing (prefill), not token generation (decode), where the harder memory-bandwidth and distributed ROCm challenges come in. So yes, impressive work, but it's not quite the same as Claude independently porting an entire inference stack to AMD GPUs.
5
2
16
6,233
Cheng retweeted
Claude API now supports changing reasoning effort mid-conversation without breaking prompt caching! Anthropic has introduced Per-message Effort (Beta), allowing developers to dynamically adjust Claude's effort level during an ongoing conversation Here's why it matters: Imagine a coding agent handling different tasks: • High effort → Analyze a complex codebase • Low effort → Summarize findings • High effort → Design and debug a fix Previously, changing effort at the request level could invalidate the prompt cache Now, developers can insert an empty system message with a new "output_config.effort" value, allowing the existing cache to be reused Key details: • No need to modify the top-level configuration • New effort level takes effect on the next user turn • Supports Claude Fable 5.2, Opus 5.5, Sonnet 5.5, Haiku 5.5, and more • Available across Claude API, Amazon Bedrock, Google Cloud, and Azure A useful upgrade for long-running AI agents, making it easier to balance reasoning depth, speed, and cost
5
1
6
751
There are basically two kinds of companies to build right now. 1. Diffusion of AI This is the "vertical SaaS" of today. There are a bunch of different models for it (FDE company, rollup, horizontal product) but ultimately it all boils down to the same thing: get AI diffused very close to where the work is being done. Profound, Cognition, Harvey etc. A lot of great companies will get built here. 2. Frontier of AI This is where the really interesting action is. Use AI to accelerate scientific progress, then build products from that progress. It is a harder category to see. But we are seeing a lot more of it at @spc: - Externa is adding hundreds of new chemical building blocks to proteins, so you can engineer a drug to match the disease instead of settling for what nature gave you. - Epana Bio is using genomics and AI to design autoimmune therapies that go after the cellular origin of the disease. - VenomicsAI is doing AI-accelerated drug discovery from venoms. - Equivariant is doing multi-physics modeling for the semiconductor industry. (There are always picks-and-shovels for AI as well. Lots of amazing companies in that category.) Category 2 is not well understood or much talked about externally. It should be because I think that is where the vast majority of model driven returns will come from.
25
28
5
309
31,006
we will soon start caring about agent portability. we'll see reverse-MCP emerge in 2027 partly out of configuration fatigue. no one likes configuring grok bot, claude and N+1 agent separately. and there's no value fragmenting what models know about you, it's worse experience for you as a user. enterprises (and likely consumers) will care about owning the specs and knowledge that the agents operate from. MCP => bring your data into the agent's env. reverse-MCP => extract your data out of the agent's env. something similar happened in the ETL space in late 2010s. ant created MCP in 2024 when it was ~1/10th the size of oai. someone who's much smaller than ant will create the reverse-MCP
> they let you connect apps via MCP but don’t offer an MCP i expect this to be a big theme in 2027
7
3
57
3,877
I’m starting to think even Anthropic doesn’t know why Opus 5.5 is so good, and are desperately trying to recreate the magic in a new, larger model. Fable 5.2 has it’s strengths, but isn’t as well rounded as Opus 5.5.
59
25
5
1,962
93,877