rama

Bio update in progress - 42%. He/I/Me
INT-8

[铁锤人 @lxfater]

卧槽，3090要可能要涨价了！！

因为老黄的刀法被人破解了

有人给3090解了个锁，同样限制在 360W，PyTorch 测试从 75 TFLOPS 跑到了 135 TFLOPS，提升约 80%！

你以为是超频，把功耗往上拉？

但他改的是显卡里控制运算速度的寄存器，这些寄存器平时有权限保护，正常改不了

他利用 GPU 启动固件的漏洞，赶在 NVIDIA 驱动加载之前拿到写入权限，把相关运算改成全速，再交给原版驱动使用

解开的主要是 FP16／BF16 输入、FP32 累加这类 AI 运算的速度限制 但测试算力多了 80%！！

4090 也有分支项目做到了：同类运算的指令吞吐，从约 165 TFLOPS 提升到 331 TFLOPS，接近翻倍

项目地址，附上👇
https://t.co/JYwB7z2PgP

[rama @bedok77]

@benjamincowen .. so bull mkt starting?

[rama @bedok77]

@ctjlewis @Joshua_Skootsky

[rama @bedok77]

@MrOngDedy so will the Indonesian Govt be suing the companies too ?

[rama @bedok77]

@satya_cosmos @tszzl 0. Same same but different from 1.

[rama @bedok77]

@bubbleboi @alexandr_wang "Altman was a former flatmate" wow , what's the odds

[rama @bedok77]

@QuintenFrancois @Bof08799061 eh.. libby tarde ?

[rama @bedok77]

@ai_for_success finally an aquarium i can manage

[rama @bedok77]

@Polymarket but can it take out the trash ?

[rama @bedok77]

@BenKoska ..er, how about Elon ?

[rama @bedok77]

@zerohedge $a16z_layoff wen ?

[Daanish Khazi @bertgodel]

We are releasing Ultramarine 1 Flash, a frontier-class model built for knowledge work. It is open-source and achieves similar performance to Fable 5, while being 60x cheaper.

It was post-trained by a team of 5 for ~$300k all-in. Available on API now, with weights releasing in 2 weeks.

[rama @bedok77]

@yacineMTB you spin me right round baby...

[𝘊𝘰𝘳𝘳𝘪𝘯𝘦 @OopsGuess]

A young Chinese neurotechnology company wiped out U.S. smart prosthetic firms that had boasted about their "advanced prosthetics" for years.

BrainCo’s bionic hand weighs just 383 grams, delivers more than 50N of grip force, interprets neural and muscle signals, and lets users grasp, type, drink, exercise and control the hand almost intuitively.

Western companies made smart prosthetics look like futuristic luxury hardware.
China is turning them into something lighter, stronger, more practical and increasingly manufacturable.

The real disruption is not that China made a better demo.
It is that China is industrializing what others kept presenting as the future.

America marketed the future.
China put it on an arm.

[rama @bedok77]

@yonann no mention of CS ?

[Rohan Paul @rohanpaul_ai]

A new UC San Diego paper builds theory of mind into a 7B agent through its training reward, so the agent learns to read a partner’s beliefs and goals and then runs on its own with no extra model at deployment.

LLM agents that negotiate, cooperate or role-play usually guess the other side's goals from prompt text or memory summaries. They keep no explicit estimate of the partner's knowledge or plans, so they can misread intent.

They train a model that infers the partner's beliefs, intents and thoughts, including what the partner thinks of the agent. That model scores candidate replies during RL, and the final agent runs alone without it.

For agents that negotiate or collaborate, train with a reward that judges replies against an inferred model of the partner's goals and beliefs.

– arxiv. org/abs/2610.12453

Title: "Mental-Models for Multi-Agent Systems"

[rama @bedok77]

@IterIntellectus @nikitabier simulation is still not good enough.. or at least they weren't up till May. https://t.co/maszmYoklN

[Nikita Bier @nikitabier]

Technical moats are mostly over at this point. If you’re building something that is meant to be durable, the only moats now are:

• Proprietary data
• Network effects
• A physical component

[rama @bedok77]

@skeptrune @bendingspoons bendingspoons acquired by Elon when ?

[rama @bedok77]

@tunahorse21 And the Witch King tipped his crown and exclaimed, "Good day M'lady!"