
@PhotoScope4all Co-Founder | Photographer 📷 Your personal AI color grading tutor. Try it free 👉 https://nitter.cf/t.co/Y77C9oxsYs
Mars
Joined November 2009
- Tweets2.3K
- Following863
- Followers55
- Likes1.5K
So after device-specific inference engines (rtx3k, v100, metal) and model-specific inference engines (qwen3/deepseek) we will see more task-specific inference engines emerging soon (like super fast decoding & prefilling on blender task only). I see those adhoc hill climbing as a positive signal showing AI is deeply integrated into daily pedestrian work and cronjobs. One can call them overfitting, but the upside of overfitting is much bigger than its downside for industry. If you are not in frontier lab, overfitting is your friend not your enemy and you should figure out a way to reach overfitting as fast as possible (via RSI etc).
Very sorry to hear this news. 😥
I never had the chance to meet Margaret Hamilton, but her contributions to the flight software that got us to the moon are legendary. I love this image of her standing next to the printouts of the Apollo code.
We’ve lost an icon. Margaret Hamilton, a computer science pioneer best known for her work at MIT leading development of the flight software for the Apollo missions to the moon, has died at 90. Read about her legacy via MIT News: news.mit.edu/2026/margaret-h…
亭山樵者 retweeted
We're excited to announce that today Python Workers are generally available (GA).
It means Python is now a first-class, fully supported language on the Cloudflare Developer Platform. Read everything about it: cfl.re/4he07Gh
亭山樵者 retweeted
Does your Claude model really need Claude Code…? 🤔
We evaluate 7 models on Claude Code, Codex, and Pi. Three surprising findings emerge:
1️⃣Harness choice has little effect on task success rate, but can significantly affect the cost
2️⃣A simple harness can be competitive
3️⃣The native harness isn’t always the best.
Millions of people are using coding agents, but the impact of harness choice remains unclear.
(1/n) More details in the thread. 🧵
AI has collapsed the cost of building products, but not the cost of knowing what to build.
- The old loop: spec, scope, build
- The new loop: build, play, design, ship
Prototypes are cheap. Great products still require taste, judgment, and a clear understanding of what users actually need.
a16z's @joshelman on the artifact a product manager actually produces: a16z.news/p/product-manageme…
亭山樵者 retweeted
DeepSeek V4.1 Flash在典型办公场景的表现怎么样?我们做了PPT生成、PPT编辑和表格分析三类测试,共37个真实办公任务;除了得分、时间和成本,还记录了模型轮次、工具调用、错误和推理Token,并分别与上一版V4、同价位的GLM-5.3-Flash,以及Opus 5、Grok 4.6进行了比较。
相比V4,V4.1比较有效地缓解了DeepSeek模型思维链过长的问题。14个表格任务中,推理Token减少38.4%,总输出减少41.7%,得分却从84.73%升至87.88%。37个任务的总时间从218分钟降至145分钟,总成本从$3.07降至$2.07,均下降约三分之一。
但它并没有明显变得更会规划。表格任务的模型轮次只从249降至235,工具调用从244降至225,仍然存在完整重算、错误后返工和成功后继续检查。它主要是每一轮想得更短,而不是用更少步骤找到正确路线。
和同价位的GLM-5.3-Flash相比,V4.1在23个PPT任务中的综合得分是86.99,GLM为82.80;耗时102分钟对171分钟。V4.1实际成本为$1.50,GLM为$0.736——V4.1消耗了更多Token、总成本约为两倍,但换来了高4.2分和快40%的结果。两者单价接近,实际表现并不是同一档。
和顶级模型相比,V4.1的执行策略仍有明显差距。14个表格任务中,V4.1与Opus 5得分接近:87.88%对87.39%;但V4.1用了235轮、225次工具调用和27.8万推理Token,Opus只用了149轮、152次工具调用和6万推理Token。PPT生成也有类似现象:V4.1比Grok 4.6高1.57个百分点,却多用了82轮和122次工具调用。只是V4.1响应更快,最终用时74分钟,Grok用了182分钟。V4.1靠吞吐和更短的单轮推理赢得了时间,并不是执行路径更短。
所以我们对V4.1的评价是:相比V4,它解决了一部分“想太久、写太多”的问题,质量、速度和成本都更实用;相比同价位模型,它愿意多做一些工作来换取更好的结果;但相比顶级模型,它在规划、工具选择和及时停止方面仍然需要更多步骤。V4.1变轻了、变快了,但还没有明显变得更聪明。
测试说明:这不是我们自己编写的一组题。PPT生成和表格分析采用OpenAI发布的GDPVal。该基准覆盖9个行业、44种职业,任务由平均拥有14年以上经验的从业者编写和审核,每项平均经过5轮专家复核。我们完整测试了全部13个PPT生成任务,并按固定规则选出14个以Excel/CSV分析及表格交付为核心的任务,不根据模型表现挑题。PPT编辑采用将发表于ECCV 2026的PPTArena;该基准包含100份真实演示文稿、2125页幻灯片和1300多项人工整理的编辑任务,我们选取覆盖文字、图片、表格、页序和样式等全部编辑类型的10项高难任务。最终结果不只看截图:所有PPTX均使用Microsoft PowerPoint原生打开并逐页渲染,再按程序检查和视觉结果核对要求;13个PPT生成任务包含622条评分项,加权满分888分,表格文件则检查分析结果、公式、结构和文件可用性。测试固定原始附件、文件校验值和任务顺序,每项任务使用独立会话;成本根据原始Usage及官方公价复算,并区分缓存命中、缓存未命中和输出Token。环境异常和无效运行留档但不计入结果。
亭山樵者 retweeted
Astra is really, really cool but I cannot current trust it for my present day engineering. At least until I have adjusted. Some random thoughts including code samples of what I recovered from my traces. lucumr.pocoo.org/2026/9/7/as…
亭山樵者 retweeted
Introducing Kitesurf: a browser built for agents, running entirely on Cloudflare Workers.
Chromium is too heavy to hand every agent one. Kitesurf is written in Rust, uses 3-7x less CPU and memory, and spins up per request.
Free in beta: blog.cloudflare.com/kitesurf…
Last-but-not-late-interaction v2! Back in 2025 when we pivoted from pairwise bi-encoder to listwise reranker, i.e. [q,d1,d2,..., dN] in one-shot. I said this is either super stupid or super smart idea, but never never mediocre. Turned out to be the pivot that opened a whole new window for rerankers in 2026.
亭山樵者 retweeted
I think there's a lot of confusion and misinformation about Zhilin's visa, H1B lottery, immigration, and why he left the U.S.
Zhilin had many opportunities to stay in the U.S. if he wanted to. In fact, I was on an email thread with a senior Apple exec (Tim Cook's reports) asking whether Zhilin would consider joining Apple. I said that he wanted to go back to his homeland. The response was, well, we have an office in Beijing if he'd like to join there.
But Zhilin was quite determined to go back and build a startup. I remember him telling me that if he didn't at least try starting his own company, he would regret it for the rest of his life. I respect that, and he was right.
Many of my international PhD students do choose to stay in the U.S. Of course the U.S. immigration process can be quite intimidating and uncertain, even for superstar PhD graduates from places like CMU.
亭山樵者 retweeted
I gave one boring PowerPoint to eight AI products. Most of them failed.
The task was as simple as it gets: "summarize the strategy in this 2-page deck."
Here's what happened:
→ Three chat assistants (Doubao, Gemini, ima) said the file was empty and gave up.
→ Three agents (ChatGPT, Cowork, Manus) eventually got it right — but only after dozens of tool calls, custom Python scripts, and in one case 520 credits burned on a single question.
→ One (WorkBuddy) ran for 20 minutes before I cancelled it.
→ remio answered correctly in under 10 seconds.
The catch? The slides were almost entirely vector diagrams stored as WMF/EMF — an old Windows format most vision models can't read. Smart enough agents can brute-force a way through, but at real cost.
This whole experiment surfaced something I've been thinking about for a while:
The bottleneck for most AI products isn't model intelligence anymore. It's whether anyone bothered to do the unglamorous work at the file layer — the WMF, the scanned PDFs, the encrypted Office docs, the audio that doesn't even look like audio.
Stronger models keep making the headlines. But the real product differentiation is happening one layer down, in the parsing pipeline nobody talks about.
Wrote up the full investigation with screenshots from all 8 products. Link: medium.com/@andrew_wy/i-test…
亭山樵者 retweeted
I've got an agent in a loop optimizing a renderer with the goal to minimize frame times (and tests to measure). It got times down from 88ms to 2ms and allocations down from ~150K to 500. Sounds good, right? Wrong. This is exactly why agent psychosis is a big fucking problem.
As an experiment, I rewrote the Ghostty core render state in Go, with access to identically laid out data structures as Ghostty and the exact same validation tests. I made a purposely naive renderer (simple, correct, but slow). 88ms per frame with 150,000 allocations (horrendous, lol)!
I then kickstarted a Ralph loop to bring the frame times down. I told it it can't modify input data structures or the public API or tests (they're correct), but it can do anything else it wants. It got to work.
It has worked for about 4 hours. I've spent around $350 on this experiment so far. The results?
88ms => 1.5ms
150K allocs => ~500 allocs
Incredible right? Nope.
My hand-written renderer I ported has frame times (same benchmark) of ~20us (0.020ms) and 0 allocations in the update path.
This is the problem with psychosis and lacking systems understanding. If you don't understand the system, you're going to accept that this is an incredible result. If you understand the system, you'll see better solutions immediately and can do roughly 75x better on throughput.
The people who blindly trust agent output are in the former camp. They're sheeple, overdrinking from a fountain of mediocrity.
Standard disclaimer: I use AI all the time. I like AI. The point I'm making is to not blindly accept results. Think. Analyze. Learn.
亭山樵者 retweeted
⚙️ Behind the build of self-improving tax agents with Codex
We co-built Tax AI with @ThriveHoldings around tax prep workflows so when reviewers fix any errors, Codex can trace the failure, improve the system, and test the change before it ships.
openai.com/index/building-se…
1/ We are sharing additional details regarding our investigation into unauthorized access to GitHub's internal repositories.
Yesterday we detected and contained a compromise of an employee device involving a poisoned VS Code extension. We removed the malicious extension version, isolated the endpoint, and began incident response immediately.
亭山樵者 retweeted
Announcing the Artificial Analysis Coding Agent Index! Our new coding agent benchmarks measure how combinations of agent harnesses and models perform on 3 leading benchmarks, token usage, cost and more
When developers use AI to code they’re choosing a model, but also pairing it with a specific harness. It makes sense to benchmark that combination to understand and compare performance.
The Artificial Analysis Coding Agent Index includes 3 leading benchmarks that represent a broad spectrum of coding agent use:
➤ SWE-Bench-Pro-Hard-AA, 150 realistic coding tasks that frontier models struggle with, sampled from Scale AI’s SWE-Bench Pro
➤ Terminal-Bench v2, 84 agentic terminal tasks from the Laude Institute and that range from system administration and cryptography to machine learning. 5 tasks were filtered due to environment incompatibility
➤ SWE-Atlas-QnA, 124 technical questions developed by Scale AI about how code behaves, root causes of issues, and more, requiring agents to explore codebases and give text answers
Analysis of results:
➤ Opus 4.7 and GPT-5.5 lead the Index: Opus 4.7 in Cursor CLI scores 61, followed closely by GPT-5.5 in Codex and Opus 4.7 in Claude Code at 60. GPT-5.5 in Cursor CLI follows at 58.
➤ Open weights models are competitive, but still trail the leaders: GLM-5.1 in Claude Code is the top open-weight result at 53, followed by Kimi K2.6 and DeepSeek V4 Pro in Claude Code at 50. These are strong results, but still meaningfully behind the top proprietary models.
➤ Gemini 3.1 Pro in Gemini CLI underperforms: Gemini 3.1 Pro in Gemini CLI scores 43, well below where Gemini 3.1 Pro sits on our Intelligence Index, highlighting that Gemini’s performance in Gemini CLI remains a relative weak spot for Google’s offering.
➤ Cost per task (API token pricing) varies >30x: Composer 2 in Cursor CLI is cheapest at $0.07/task, followed by DeepSeek V4 Pro in Claude Code at $0.35/task and Kimi K2.6 in Claude Code at $0.76/task. At the high end, GPT-5.5 in Codex costs $2.21/task, while GLM-5.1 in Claude Code costs $2.26/task. For both models this was contributed to by high token usage, and in GPT-5.5’s case by a relatively higher per token cost.
➤ Token usage varies >3x: GLM-5.1 in Claude Code uses the most tokens at 4.8M/task, followed by Kimi K2.6 at 3.7M/task and DeepSeek V4 Pro at 3.5M/task. GPT-5.5 in Codex uses 2.8M tokens/task, substantially more than Opus 4.7 in Claude Code at 1.7M/task. In GLM-5.1’s case, higher token usage, cost and execution time were partly driven by the model entering loops on some tasks.
➤ Cache hit rates remain high but vary materially: Cache hit rates range from 80% to 96% across combinations. Provider routing, harness prompt structure and cache behavior can materially change the economics of running the same model given cached inputs are typically <50% the API price of regular input tokens.
➤ Time per task varies >7x: Opus 4.7 in Claude Code is fastest at ~6 minutes/task, while Kimi K2.6 in Claude Code is slowest at ~40 minutes/task. This is contributed to by differences in average turns per task, token usage and API serving speed. Opus 4.7 had materially lower amount of turns to complete a task than all other models while Kimi K2.6 had the most.
➤ Cursor made real progress with Composer 2: Composer 2 in Cursor CLI scores 48, near the leading open-weight model results, while being the cheapest combination measured at $0.07/task. Cursor has stated Composer 2 is built from Kimi K2.5, showcasing they have made substantial post-training gains.
This is just the start. We are planning to add additional agents (both harnesses and models). Let us know what you would like to see added next.
亭山樵者 retweeted
I think this is bigger than it sounds at first glance.
Thinking Machines hasn't just unveiled "ChatGPT, but better." Instead, they've introduced something that addresses a much deeper issue: the very way we interact with AI.
So far, AI often feels like email with very clever replies. I say something. Then the model waits. Then it replies. Then I wait.
Thinking Machines' new Interaction Model attempts to break down precisely this barrier.
It can simultaneously listen, see, speak, interrupt, react, think in the background, and use tools. Not as a cobbled-together pipeline of speech-to-text, turn detection, and agent hacks, but as a native model capability!
Good collaboration doesn't happen because someone gives a perfect answer in the end. It happens because someone is present in the moment.
If this works, AI shifts from "prompt in, answer out" to something that feels more like collaborative work. A model that notices when you hesitate. That jumps in when it sees something. That anticipates your next move while you speak. That not only gets smarter, but also better at maintaining a flow of conversation with people.
ngl really impressed by their examples.
People talk, listen, watch, think, and collaborate at the same time, in real time. We've designed an AI that works with people the same way.
We share our approach, early results, and a quick look at our model in action.
thinkingmachines.ai/blog/int…
亭山樵者 retweeted
Hard to believe it’s been 10 years since AlphaGo! It was wonderful to catch up with Lee Sae Dol last week in Korea and join Shin Jin-seo for a special Go match. Great to reminisce about AlphaGo & super interesting to hear how it changed the way players approach the game of Go!
#AlphaGo WINS!!!! We landed it on the moon. So proud of the team!! Respect to the amazing Lee Sedol too




































