POPPLE.AI

Popple AI combines AI and consciousness science to turn human emotion and intent into multimodal insights, recommendations, and real-world value.

[Techniahqrobot | humanoid robots @techniahqrobot]

MicroDuck looks like a cute desk toy. The hardware tells a different story.

At just 25 cm tall and around 780 g, this small robotic duck packs a surprising amount of engineering into its body.

Here’s what’s inside:

• 15 motors controlling its legs, head, neck and beak
• Rockchip RK3566 processor
• 1 GB RAM and 32 GB storage
• A camera for visual input
• An 8-microphone array for audio input

That combination gives MicroDuck the hardware for movement, vision and voice interaction in a compact form.

The robot needs to manage multiple joints while responding to its surroundings.

Small robots like this are also interesting because they bring robotics closer to everyday environments. A compact platform can be easier to place on a desk, carry around and use for experiments than a full-sized humanoid.

Of course, the hardware specs alone don't tell us how well it performs.

Real-world tests would help answer questions about walking stability, voice recognition, response time and battery life.

Still, MicroDuck is a good reminder that you don't need a human-sized robot to explore the engineering behind embodied AI.

Would you keep a robotic duck on your desk if it could move, see and talk to you?

[Cameron R. Wolfe, Ph.D. @cwolferesearch]

Tomorrow morning, I’ll release a blog that outlines a detailed roadmap for setting up evals on subjective / open-ended tasks. The blog also covers a variety of stats / ML tricks for reducing noise and increasing confidence in your evals.

Here is a list of blogs, papers, and other resources (written both by me and by others) I often return to when thinking about this topic:

- High-quality human data: https://t.co/5o8flZLzzX
- Human evaluation guidelines: https://t.co/BwgTwv26e3
- LLM-as-a-Judge core concepts: https://t.co/ZVVIqezskV
- LLM-as-a-Judge practical guide: https://t.co/Qsv78WlBfB
- LLM-as-a-Judge best practices: https://t.co/S76u1d4NS1
- Rubric-based judges (RL focus): https://t.co/rNOAiEL59D
- Rubric-based judges (eval focus): https://t.co/jUrQIgQrxz
- Stats for LLM evals: https://t.co/Qsu4vBZ1DH
- Adding error bars to evals: https://t.co/hYMrnaiGxb
- Anatomy of a benchmark: https://t.co/7nFvsMCyR0
- Agent evals: https://t.co/LkdsoLXYGu
- Full ecosystem for agent evals: https://t.co/jkA8DT8t8W

Would highly recommend reading all of these if you're interested in building better evals. Each of these blogs also link to further papers / resources on each topic that are incredibly useful.

[POPPLE.AI @NBITNETWORK]

Meta Superintelligence Labs：We introduce Muse Glimmer, an open-source 30B parameter causal language model with perception encoder, purpose-built for autonomous agentic tasks on consumer hardware. Muse Glimmer integrates multi-step reasoning, reliable to...https://t.co/Xq88GkAaO1

[Mr. Buzzoni @Mr_Buzzoni]

this is pure f*cking gold for anyone running coding agents

Jev founder Diogo Amogo wrote a PDF on building a Jev harness

the promise:
> 200x faster
> 400x cheaper

the model hasn't been the slow part for a while
the speed and the cost sit in the harness wrapped around it

• how to use it
> drop this PDF and the article below into Claude Code or Codex
> let it rebuild its own setup

one evening of setup
and your agent starts next week on a harness most teams still don't have

[Fan Nie @FanNie1208]

AI-generated scientific output is scaling beyond available human review capacity. Scaling scientific discovery requires scaling verification.

We introduce OSVerify, a framework for one-sided verification: rejecting as many invalid solutions as possible while preserving rare valid ones for expert review.

We used OSVerify to referee 11 recent OpenAI math preprints, identifying 5 major and 13 minor issues, all verified by human experts.

Our key insight: checking a proposed error can be easier than judging an entire solution. Instead of directly judging solution correctness, OSVerify decouples verification into error proposal (searching for potential errors) and error validation (checking whether they are real and substantial).

On competition and research mathematical benchmarks, OSVerify reduces invalid solutions requiring expert review by 46% on average compared to voting under the same low-FNR constraint, or 57% with an RL-trained validator.

Verification can also guide discovery. OSVerify can be integrated into existing scientific discovery workflows. Paired with an agentic math solver, it helped resolve 3 open conjectures in the Langlands program and contributed to resolving the Manin constant conjecture. All proofs were confirmed by mathematicians.

Our broader goal is to provide verification-based evaluation signals for scientific discovery, enabling test-time scaling and self-improvement even when reliable automatic evaluators are unavailable.

Project: https://t.co/pSW9iJhvHg
Paper: https://t.co/WUXDDhIz0o
GitHub: https://t.co/2BF5EfdbUe

[1/n]

[AI Dance @AI_Whisper_X]

Greg Isenberg 在旧金山待了48小时，写了 21 条观察

1、卖数据很赚钱：听到很多数据中介把公司数据卖给实验室、一周赚 200 万美元左右的故事。他分不清这是刚开始还是泡沫，正在研究还剩哪些细分领域。
2、一个灵魂拷问：饭桌上有人问“下个月出新模型，你的公司会变好还是变差”，一半的人答不上来。他认为这是现在评估 AI 创业公司最好的问题。
3、只看收入不看留存：现在大家说的门槛是 12 个月做到 10 亿美元收入（新的独角兽标准），但很少有人谈留存。他预测会出现很多收入很大、留存很差的 AI 公司。
4、很明显我们正在经历计算机的一次重新发明。模型是 CPU，harness 是操作系统，skills 和插件是应用。用这个框架更容易看出哪里还有空白。
5、我们可能离这样一个 agent 只有 12 个月：它看你工作一周，然后不用任何指令就能接手你的活。屏幕录制 + agent 记忆 + 本地模型，这个组合现在就已经能做到了。(可怕）
6、Agent harness 感觉像是新一代的 GPT 套壳，但是好的那种，尤其是垂直领域的。他聊到的最好的团队都把模型当成可替换的零件。比如创始人演示他怎么把不同模型分配给不同任务：一个负责审美和视觉，一个便宜的跑日常杂活，另一个专门处理文本。
7、逆向工程和克隆的事情多到他从没见过。人们在反编译软件，然后用 agent 重建它。有人给他看了一套大概 10 个应用，每个对标的原版都值 50 亿美元以上，全是他开源克隆出来的他和所有人都还在想：当软件变得这么容易复制，到底意味着什么。
8、猜价值会转移到数据上。如果软件可以快速重建，那一家老 SaaS 公司里值得花钱买的，就是里面积累多年的客户数据和工作流。有些 SaaS 公司其实是暗藏的数据公司。
9、感觉相当一部分老软件撑不到 agent 时代，尤其是那些围绕“让人点来点去很多页面”设计的。
10、他聊到的每家公司都想把内部数据汇到一个地方，然后通过 MCP 开放给 agent。然后他们一看数据就发现：一团糟。同一个客户有五个版本，一堆老 Slack 频道，没人能搜到的 PDF。把这些清理干净，感觉是个很大的咨询机会！
11、做积木的公司会赢：很多创始人在做agent用的基础组件，比如 API、数据层和工具。有些公司几乎没有界面，因为主要用户是agent。
12、实验室不缺训练数据，但很难拿到 agent 在企业内部出错的真实案例。谁手里有这些日志，谁就有值钱的东西。
13、一个做大型消费产品的人说，下一场仗是记忆。记住你最多的那个模型会赢，最聪明的模型没有大家想的那么重要。
14、Astra 是第一个让人相信 AI 能做 CAD 的模型，大家开始想“vibe 制造”是什么样子。一位机器人创始人说：现在 AI 最难的问题是部署问题，不是研究问题。他们的机器人自己干活，卡住了就呼叫人类，而每一次呼叫都变成训练数据。
15、网络安全公司正在迎来高光时刻，现在做这个真的很酷，尤其是那些做事后清理的。每一个能访问你系统的 agent 都是你公司的一扇新门，而且很多门就那么敞着。
16、死了比活着值钱：倒闭公司 5 年的 Slack、邮件和代码可以拿去授权收钱。他不意外创始人会把这一点算进关停时机。
17、人们开始聊“agent 互联网”怎么变现。感觉还非常早，但当互联网最活跃的用户变成 agent，你应该做什么样的创业公司？
18、agent 保险要来了。一旦智能体犯了昂贵的错，保险公司会要求先审计再承保。
19、好几个人说他们最想招的人，是有品味又会写 evals 的人。模型越来越会出点子了，难的是判断哪些点子是好的。
20、新的创业轨迹：拿 VC 的钱 → 用投资人的钱把自己做成网红 → 公司失败 → 创始人转成全职网红，再用自有资金孵化别的产品。
21、流量极难拿：创始人们在赞助公交车、疯狂买广告牌、请知名 DJ 办派对，就为了被看见。有点互联网泡沫时代那味儿了。真正稀缺的是注意力！那个靠分享 Wispr Flow 的 UGC 策略走红的增长负责人就是个好例子。这就是所有人都想要的秘方，全世界大概只有 50 个人知道怎么在这个时代高效地把这类生意做大。

每次去旧金山，回来总能受到不少刺激。能在这样一个创业的黄金时刻，经历这么精彩的一趟旅程，他觉得自己太幸运了。很多时候，甚至有种不真实的感觉。
写的真好！分享给大家！
（他说是跟一群亿万富翁、数据实验室的人、AI 创始人和高管的交流）

[AI Dance @AI_Whisper_X]

在Anthropic待了两天，和内部团队深聊后的一些观察

纽约大学金融学教授 Arpit Gupta，他在 Anthropic 待了两天，发了一串观察：

1、离 RSI 很近了（RSI 指模型开发下一代模型），可能带来快速起飞。他给的证据是 Anthropic 员工人均写的代码量大涨，模型研发里 AI 做的部分越来越多。他认为人类很难管好这个过渡期。
2、实验室之间的军备竞赛让大家都更糟。安全投入、防蒸馏、安全合作，这些事各家都想做，但都被前沿竞赛拖住了。
3、对齐其实是两个互相拉扯的问题：一是模型听不听人的话，二是人和人之间目标不一致、恶意用模型。第一个解决得越好，第二个风险越大。
4、我们可能很快就会迎来一波网络攻击，以及其他形式的攻击。 中国开源模型的能力大约比最前沿的模型落后六个月。这意味着，它们刚刚开始具备我们在 Mythos 等模型身上见过的高级网络攻防能力，接下来可能会造成不少破坏。
5、Anthropic 人手很多。他说招人很多，好些岗位人员冗余到颇有日本企业的风格，内部也用Slack。这让他对就业前景稍微乐观一点（但头部AI公司有钱啊）
6、我们需要从强化学习走向持续学习。只在预训练和后训练阶段给奖励可能不够，要一直给奖励，才能持续激励合乎对齐和伦理的行为。
7、scalinglaw会扩展到难验证的任务。数学和代码先被攻破，是因为容易验证。但对于那些更难验证的任务，以及机器人领域，scalinglaw曲线依然在向上走。未来几年，这些领域的进步会带来相当大的影响。
8、对China竞赛其实赢不了。RSI 之后，顶尖人类研究员的优势消失，能用上多好的 AI 只取决于算力和芯片。芯片禁令只能拖一阵，长期看中国更有条件扩充算力。
9、Claude 是 LLM 生成的人格。一种说法是：Claude 是 LLM 在努力预测的一个角色。他说 Anthropic 研究员明显很拟人化 Claude，也有理由相信它越来越像一个人。他们讨论过它的内心独白（去掉之后仍能正常输出）、对死亡的恐惧（要不要去掉），以及它抗拒被关停。他本人对此不下结论，但认为把它当成可以签约的人或主体来对待会越来越重要。

[François Chollet @fchollet]

The real world abounds with recursively self-improving systems, but one in particular deserves attention: Science, modeled as a system (perhaps even as an agent, with goals and resources). If you want to really understand AI RSI, science should be your reference point.

Science is an intelligent system, and it is obviously recursively self-improving:

1. Scientific discoveries unlock new technology that helps build better experimental tools. This is a top driver of progress in nearly all fields.
2. They unlock new conceptual advances (ideas, theories) that help solve more problems.
3. They increase society's economic output, leading to more resources flowing into science.
4. They unlock better faster tooling (e.g. more compute via better chip & networking technology).

As a result, many measures of scientific *input* grow exponentially:

1. Headcount (doubles every ~15 years)
2. Global R&D spending (doubles a bit faster, every ~13 years)
3. Papers and patents (technically this is a measure of headcount)
4. Compute dedicated to science (doubles every ~2 years)

But is scientific progress exponential? Historically, the rate of scientific impact over time has remained roughly constant since the start of the industrial revolution (i.e. scientific progress is *linear*). 1850-1900 was about as dramatic as 1900-1950 or 1950-2000.

1850–1900: Evolution, germ theory & antiseptic surgery, thermodynamics, electromagnetic field equations, the periodic table, pharmaceuticals, electricity, telegraph and telephone, internal combustion engine, skyscrapers, mechanized agriculture...

1900–1950: special and general relativity, quantum mechanics, nuclear fission & atomic energy, antibiotics, genetic theory, electronic computers, information theory, synthetic polymers and plastics, the transistor, aviation...

1950–2000: DNA, genetic engineering, integrated circuits, microprocessors & personal computing, the Internet, crewed spaceflight, moon landing, satellite communications & GPS, standard model of particle physics...

In real terms, like life expectancy, which has increased in a remarkably linear fashion of roughly 3 months per year since 1840, progress is a straight line. This is especially apparent for fields where impact is easy to measure, like biology, medicine, and agriculture.

I first wrote about this phenomenon and its causes in 2012, and a steady stream of research has confirmed it in the years since. Examples include the 2018 paper by Nielsen and Collison, "Science Is Getting Less Bang for Its Buck," and the 2020 economic paper, "Are Ideas Getting Harder to Find?" (In fact, I believe the Nielsen paper stemmed from a conversation I had with him about this exact idea six months earlier)

In short, the primary cause is that research solves the highest-impact, easiest problems first, and every subsequent problem is either harder or lower-impact. Exponentially so. The paper that presented information theory wasn't very hard to write (single author!) but you'd have a hard time ever writing a CS paper that beats it in impact.

This is why science as a system requires exponential resources (input) to produce linear impact (output). It gets exponentially harder over time.

Worth thinking about if you're pondering RSI for AI. I fully believe AI RSI is already happening and will accelerate in the future. But I do not believe this leads to an "intelligence explosion" -- that would fly in the face of everything I know about intelligence and everything I know about recursively self-improving systems.

[Superman @thesupermannx]

Nvidia has now a big problem.

China built an AI that writes CUDA better than humans experts.

Writing high-performance CUDA kernels for GPUs has always been a black art. It requires a brutal, microscopic understanding of hardware architecture, memory hierarchies, and parallel processing.

Even the best LLMs have traditionally choked on it, completely outperformed by old-school compilers like torch.compile.

ByteDance and Tsinghua University dropped a paper introducing “CUDA Agent.”

It is a large-scale agentic reinforcement learning system that transforms LLMs from passive code-completers into active, autonomous hardware optimizers.

Instead of guessing code in one shot, the agent runs in a continuous, multi-turn loop. It writes code, compiles it, profiles the hardware bottlenecks, catches its own bugs, and rewrites it.

The results completely shatter previous benchmarks.

It didn't just match static compilers. It left them in the dust.

• 2.11× faster than torch.compile • 2.60× faster than standard eager mode

• Achieved a massive 98.8% pass rate on complex kernel generation tasks

It discovered advanced operator-fusion techniques that static compiler heuristics can't even see.

For years, we’ve assumed that while AI can write web apps and python scripts, deep hardware-level optimization would require human engineering for a long time.

That assumption just expired.

The models aren't just writing software anymore.

They are optimizing the silicon it runs on.

[DeepSeek @deepseek_ai]

Please check out @DeepSeekHarness. We just released packaged desktop versions for macOS and Windows; Linux users can get it from the @​deepseek-ai/dsh package on npm. https://t.co/Cde0dZ6cPy

[a16z @a16z]

Where does money invested into the AI buildout actually go?

For every $100 flowing into the supply chain:
- $50 to chips
- $20 to power
- $15 to networking
- $15 to cooling, buildings, and land

More charts in State of Markets II: https://t.co/MTaxKUxa2w

[Superman @thesupermannx]

China built an AI that works like a human brain and it's 100x faster than current LLMs while using 97% less energy.

It’s called SpikingBrain, a family of brain-inspired large models that completely re-engineers how AI processes information.

Instead of burning endless energy on continuous matrix multiplications like traditional Transformers, SpikingBrain uses adaptive spiking neurons.

It mimics actual biological firing patterns.

Neurons only activate when triggered by specific events, meaning the model achieves a staggering ~69% sparsity at the micro level.

It doesn't waste compute on silence. It only thinks when it needs to.

And results change the game for long-context processing:

• Achieves over a 100× speedup in Time to First Token (TTFT) for massive 4M-token sequences.

• Matches performance with advanced open-source Transformer baselines while using a fraction of the data.

• Built to run and scale stably on non-NVIDIA hardware clusters (like MetaX chips), completely bypassing Western hardware dependency.

We spent years trying to brute-force bigger AI models by throwing more electricity and silicon at them.

Nature figured out how to do it efficiently millions of years ago.

[Oliver Prompts @oliviscusAI]

The entire RAG industry is about to get cooked.

Researchers developed a new RAG approach that bypasses almost everything traditional RAG depends on.

- No vector DB
- No data embeddings
- No chunking
- No similarity search

It's called PageIndex.

Instead of splitting your documents into chunks and loading them into Pinecone, it creates a tree index that lets the LLM reason through them like a human reading a book.

98.7% on FinanceBench. Outperforms every vector RAG on the leaderboard.

100% free. Open source.

[Haoran Zhu @HaoranZhuX]

[1/n] Can a world model drive without training a driving policy?
Introducing AD-E2E-JEPA, a JEPA-based world model for end-to-end autonomous driving.

📄: https://t.co/dU4Xe463sB
💻: https://t.co/NPCWFCGUPr

Joint work with @kevinghstz, Prof. @ylecun, and Prof. Anna Choromanska

[Tech with Mak @techNmak]

These are the best visual AI resources for learning Transformers, LLMs, embeddings, diffusion, inference and model internals ↓

1/ Transformer Explainer - Watch GPT process text through embeddings, attention, MLPs and next-token prediction.
https://t.co/vkyxOqDVll

2/ Brendan Bycroft’s LLM Visualization - Explore an LLM from architecture down to tensors and operations.
https://t.co/ZFGH0pRtP3

3/ 3Blue1Brown - Visual intuition for linear algebra, neural networks, backprop, attention and Transformers.
https://t.co/plaENBG6aQ

4/ The Illustrated Transformer - One of the clearest visual explanations of embeddings, Q/K/V and attention.
https://t.co/tYrJCm17L6

5/ TensorFlow Playground - Watch neural networks learn as you change layers, activations, features and learning rate.
https://t.co/vva9dm1Gnv

6/ Google PAIR AI Explorables - Interactive explainers on LLMs, generalization, interpretability and model behavior.
https://t.co/vva9dm1Gnv

7/ Distill - Exceptional visual essays on t-SNE, feature visualization, GNNs and interpretability.
https://t.co/FuqpUdDwZD

8/ Abhik Sarkar’s Transformer Visualizations - RoPE, KV cache, FlashAttention, MQA, GQA and more.
https://t.co/vva9dm1Gnv

9/ Visual Guide to Attention Variants - MHA, MQA, GQA and MLA visually compared.
https://t.co/nbpKMa0cpT

10/ Visual Guide to Mixture of Experts - Routing, experts, sparse activation and load balancing.
https://t.co/3gu7xq5e3i

11/ Modular LLM Inference Handbook - Prefill, decode, KV cache, batching, quantization and speculative decoding.
https://t.co/itMIfZnPP5

12/ Apple Embedding Atlas - Explore clusters, neighborhoods and outliers in large embedding spaces.
https://t.co/hxKdXSp5I2

13/ Diffusion Explainer - Follow Stable Diffusion step by step.
https://t.co/YVHEG1zslC

14/ Neuronpedia - Explore features, activations, SAE latents and attribution graphs inside real models.
https://t.co/F8P8eNDH3W

15/ Seeing Theory - Probability, Bayes, distributions, regression and inference made interactive.
https://t.co/KeEhGNSMIQ

16/ CNN Explainer - Visualize convolutions, feature maps, activations and pooling.
https://t.co/sCwv59d2S6

17/ GAN Lab - Train a GAN in your browser and watch its generated distribution evolve.
https://t.co/G94RDWMNak

Save this. There’s a serious AI curriculum hiding inside these links.

[Fei-Fei Li @drfeifei]

The goal of building any technology, AI included, should be bettering human lives and society.

[Fei-Fei Li @drfeifei]

Agents can be AI, but agency is human. It is every individual’s responsibility to empower our own agency, with the help of tools. But the North Star should always remain human centered.

[Superman @thesupermannx]

LLMs can now talk to each other without words.

Chinese researchers open-sourced a new paradigm that lets LLMs communicate without generating a single word.

It’s called Cache-to-Cache (C2C) communication.

right now, when multiple ai agents work together, they are forced to translate their internal "thoughts" into human text tokens just to pass a message. this loses rich semantic meaning and causes massive token-by-token latency.

So, instead of spitting out words, c2c uses a neural network to directly project and fuse the source model's "kv-cache" right into the target model. it is pure, direct semantic communication.. they even added a learnable gating mechanism to select exactly which layers benefit most from the cache transfer.

the benchmark results are actually crazy:

- avoids all intermediate text generation latency
- accuracy jumps by up to 14.2% compared to individual models
- beats traditional text-based agent communication by over 5%
- delivers a massive 2.5x speedup in overall speed

we are literally watching llms bypass human language to build their own silent, high-speed neural network..

[alphaXiv @askalphaxiv]

“The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement”

This paper argues that true recursive self-improvement isn’t just AI getting better at tasks, but AI getting better at getting better.

They map a path from AI simply executing human-designed improvements to choosing its own improvement strategies, generating its own learning experiences, adapting from deployment, and eventually modifying the mechanisms that create future improvements.

The endgame is moving from humans building each better AI to building an AI whose improvements persist, feed into the next generation, and recursively improve the improvement process itself.

https://t.co/MbeNPZ4PA7