_Harmony_ retweeted
🤯RSI突破！腾讯AI混元四步炼出终端Agent神器，Qwen3.8-27B性能直接从51.7%提升至86.5%，小模型硬刚一线！
腾讯混元最新论文Terminal-RSI正式开源递归自我改进（RSI）打法：交替两轮“成功轨迹自重写 + 策略精炼”，让27B模型在终端任务上实现质变。
要点速览：
🔹方法极简却狠：Base → RSR-1（验证成功轨迹Recursive Self-Rewrite做SFT）→ RPR-1（继承critic的RL精炼）→ RSR-2 → RPR-2，两轮闭环，前一轮模型直接生成下一轮训练数据。
🔹成绩炸裂：Terminal-Bench 2.1从51.70%飙到86.52%（+34.82pp）；TB 4.0从0%→19.84%；TB Hard达79%；多基准平均大幅跃升约30个百分点。
🔹长程也破防：Long-Horizon Terminal-Bench process reward 0.21→0.37，严格完成0/46→5/46；SWR100成功率直接翻倍。
🔹技术硬核：验证轨迹重写 + critic权重继承校准 + TB-GAE信用分配 + Success-Prioritized Reward，把“部分进展”真正转化为“严格验证通过”。
🔹意义拉满：延续T1/RSR系列，证明RSI在终端Agent上可稳定爬坡，27B小模型就能冲到86%+水平，逼近甚至超越部分更大模型。
查看论文：arxiv.org/abs/2610.22192
这不是概念炒作，是真正把验证过的成功经验反复炼成能力，终端Agent又要卷起来了！
#RSI #递归自我改进 #腾讯混元 #Qwen #TerminalBench #AIAgent #大模型 #TerminalRSI #AgentRL #混元AI
💥太全面了吧！RSI资源地图来了！开源神器 Awesome RSI 资料库正式整理完成！这是一份精选的 GitHub 研究合集，专为探索递归自我改进的开发者与研究者打造！
RSI 不是普通迭代，而是 AI系统不仅提升自身能力，还能改进“产生改进的机制”本身。从自训练、记忆演化、Harness优化，到自动化AI研发、具身系统、安全对齐，全都系统梳理好了，再也不用一篇篇论文自己拼图。
核心亮点
🔹先打基础：奠基论文、形式化分析与综述一网打尽
🔹术语清晰：区分自我精炼 / 持续自改进 / 递归自改进 / RSI基底
🔹分层拆解：模型层、Harness层、多智能体、自演化框架全覆盖
🔹评测路径完整：直接自改进评测 + AI研发/智能体能力基准
🔹社区驱动、CC0开源，持续更新中（已600+ Star，增长飞快）
RSI资源主要分类
1️⃣基础与入门 — 奠基论文与术语厘清
2️⃣模型层RSI — 自训练、自奖励、合成数据、自博弈
3️⃣Harness层RSI — 提示优化、记忆演化、脚手架演化、自演化智能体
4️⃣多智能体自改进 — 协同演化与辩论机制
5️⃣编码/软件工程自改进 — 自修改编码Agent
6️⃣自动化AI研发
7️⃣具身与物理自改进
8️⃣进化与u开放式RSI
9️⃣安全、对齐与理论
🔟基准与评测 + 框架工具合集
非常适合做RSI研究和从事AI的人系统学习，快速建立全局认知。
👉 直达链接：github.com/lobehub/awesome-r…
#大模型 #算法 #人工智能发展 #AINEXT #RSI #RecursiveSelfImprovement #AwesomeList #开源AI #AI智能体 #AI安全
_Harmony_ retweeted
🚀 Today's DeepSeek Harness 0.2.1 release supports the new DeepSeek-V4.1-Flash-1017 model. The new model was specifically optimized for @DeepSeekHarness during training, improving task performance while shortening task runtime and reducing output usage and token costs. To take full advantage of the new model’s capabilities, please upgrade to DeepSeek Harness 0.2.1-rc.1 or later.
🚀 DeepSeek-V4.1-Flash-1017 is here!
Stronger benchmark performance, faster task completion, and shorter outputs than the previous version—a Pareto improvement in performance and cost. Now live on our website, app, and API.
🛠️ Optimized for @DeepSeekHarness, Tencent CodeBuddy, WorkBuddy, and more. Upgrade to DeepSeek Harness v0.2.1-rc.1 or later for the best experience with the new model.
1/4
_Harmony_ retweeted
Wow🙀🙀 DeepSeek V4.1 Flash 1017 joins BuildingBench’s top five—an exciting leap for open models!! 🔥
We tested its 3D building reconstruction from real photos. An open model delivering this level of spatial reasoning at this cost is exciting to see! 👀
🎈 81.9 overall, up from 72.9 for the previous version. A 9-point jump!
🎈 Slightly higher than Fable 5.1’s 81.4, and essentially level with Sonnet 5.5’s 82.1.
🎈 Estimated median cost: $3.01/task. About 1/11 of Fable 5.1’s cost and 1/15 of Sonnet 5.5’s.
_Harmony_ retweeted
It's so hilarious that so many people THINK they saw a road-ready, fully equipped car with real road tires, batteries, interior and everything, just flying across a stage.
That's what Elon wanted you to see. But the car was just a thin shell over a lightweight chassis with thrusters. The tires weren't real road tires (those would have been too heavy). There were no seats inside. No long-range batteries (too heavy). The car wasn't real. It was simply a MODEL of a car. (Notice they never showed you inside the car, right?)
But people are dumb. And easily mesmerized by showmanship. So mission accomplished. Elon managed to convince low-IQ people that Tesla has flying cars nearly ready to start selling to the public (yes, people actually believe that). These are the same idiots who also think they're going to receive a "Universal High Income" because their government loves them.
LMAO. The dumb ones are already being culled. With jabs and flu shots, of course. Because you have to be cognitively retarded to be taking those in the first place.
_Harmony_ retweeted
This is what Tesla’s value (and Elon Musk’s net worth) is based on:
Expensive gimmicks that provide zero value in the real world. This car has a 10K PSI nitrogen tank attached to it.
It’s basically a flying bomb. It can never be sold, and it has zero practical use.
Elon was always a fraud. But the stupid shareholders will be impressed by this worthless stunt, and they’ll buy even more shares of Tesla, making Elon even richer.
_Harmony_ retweeted
We gave DeepSeek’s updated V4.1-Flash-1017 the same 100 pull requests with known vulnerabilities and an hour per PR to find them in the @BugBunny_ai VulnPR-100 benchmark.
It found 54 for just $7.53, making it the best-performing open-weight model on our leaderboard.
That matches Grok 4.6’s 54 findings at roughly 16× lower review cost. It also finishes only three findings behind Gemini 4 Argon while costing roughly 22× less.
The previous DeepSeek V4.1 Flash found 44 for $12.79. The update catches 10 more vulnerabilities while costing 41% less, bringing the cost down to just $0.14 per vulnerability found.
Better recall, lower cost, and the strongest open-weight result we’ve measured. Seriously impressive work from the @deepseek_ai team.
_Harmony_ retweeted
They're talking about Joshua Brown
who was one of biggest $TSLA fanbois
He post numerous videos showing off his "hands free" driving with Tesla autonomous tech
It cost him his life and destroyed his family.
Tesla and Elon knew $TSLA autopilot / FSD couldn't tell difference between bridge and big rig but still released the software 🤡
A 2 yr old knows the difference.
Josh Brown's head was sliced off bc he trusted Elon.
Elon prioritizes money over the lives of his own fanbois 🤡💰
_Harmony_ retweeted
GPT 6.1 Sol Ultra vs DeepSeek V4.1 Flash 1017 Max, I still can't believe a model that's 90% cheaper can produce results this good, how is this even possible? Both models took roughly the same amount of time since most of it was spent rendering, but the quality DeepSeek managed to achieve is incredible, especially considering the API cost was just 0.44 cents for the entire test! I think having access to free web search is making a huge difference and giving DeepSeek a much bigger advantage than before, especially with its native multimodal capabilities. The more I test this model, the more I wonder how they're managing to offer so much at such a low price, I'm still running even bigger tests with different models to see how far it can go, more results coming soon!
_Harmony_ retweeted
先ほど公開されたDeepSeek-V4.1-Flash-1017のベンチマークスコアが完全におかしいことになっている（アクティブパラメータが8B、16Bで、GPT-6とOpus5、Fableにパフォーマンス、コストともに圧勝）
元々アメリカのフロンティアAIラボが荒れていたタイミングだったので、DeepSeekショック再びの可能性
_Harmony_ retweeted
There is this persistent myth that Chinese EV companies “copied” Tesla.
Nuffin can be further from the truf. There were multiple EVs in China’s market every bit as advanced as the Model 3 and Model S (NIO, Xpeng, Aeon). Elon Musk threw in the towel in 2018/19 when he realised that only China can manufacture EVs.
This was true then and is true today.
Multiple engineering teams from Tesla Shanghai have been sent to the company’s US factories (Fremont/Austin) to get them in order… they all failed necessitating the +100% tariff.
Tesla Shanghai saved Elon’s ass. Exports from Shanghai are one of two sources of company profits (the other one being long suffering Americans forced to pay inflated prices in a protected market).
The Roadster stunt is an expression of a Western concept called “face.”
When they can’t do anything real – like manufacture and sell a roadster – Westerners resort to distractions and party tricks.
Some play with compressed air remote controlled models and others talk about how their products might end the world. It’s really quite weird.
wsj.com/business/autos/what-…
_Harmony_ retweeted
This may not be a popular take, but it seems like Tesla made the same mistakes with the Roadster as they did with the Cybertruck.
They went completely overboard.
While the Roadster demo was spectacular. It was just a demo and most owners won't get or use the features shown off.
The thrusters dictate so much of the vehicle's design and likely single-handedly delayed the vehicle. The thrusters only provide a one-time, 5 second boost before the tank is depleted. From estimates gathered online, the tank also costs $500 - $2,000 to fill.
Instead of the original sleek design that was already prototyped with 2+2 seating and could be used around town, Tesla designed the vehicle around the thrusters. It's not practical and you have to mount on a windshield wiper if it starts raining.
While the demo was amazing, this isn't the vehicle most people wanted and Tesla likely spent 10s of millions of dollars developing the thruster system instead of focusing on releasing a quad-motor version of the 2017 Roadster on time.
Is this what you wanted in the Roadster?
_Harmony_ retweeted
Guy who said we would have coast to coast self driving cars in 2019 and 1 million robotaxis in 2020 says advertisers aren't allowed to overstate claims anymore. 😂😂😂😂😂🤣🤣 LOL 😆 😂 🤣 😅 😄
_Harmony_ retweeted
Alex Gibney: “A lot of these deaths related to Full Self Driving or Autopilot, 67 so far. If that had happened at GM over a brake malfunction, it would be massive congressional hearings. A lot of those things were settled so the actual facts of the matter wouldn’t come to light. In the film, we have the case of one which is a pedestrian who was killed by a runaway Tesla with malfunctioning Autopilot. They took that to trial and won a rather massive judgement and the evidence is overwhelming not only how bad the system was, but also how hard Tesla tried to hide the evidence of malfeasance. In a lot of cases, Autopilot has a function where milliseconds before the car slams into a wall, Autopilot will turn off, and so Tesla can say with a straight face Autopilot was not engaged at the time of the crash”