RDM | AIE @ https://xitter.gq/t.co/terMAcKmHy | Resident @primeintellect | Founder @polySpectra | Cohort 1 @activatefellows @berkeleylab | PhD @caltech | AB @princeton | #rwri
New Haven
Joined December 2013
- Tweets4.3K
- Following3.8K
- Followers3K
- Likes13.5K
Pinned Tweet
“Study hard what interests you the most in the most undisciplined, irreverent and original manner possible.”
― Richard Feynmann
I am currently studying:
ralph wiggum, rlms, dspy, gastown & related agent orchestrators, agent-native apps and orgs
How about you?
Raymond Weitekamp retweeted
Imagine MetaSynth, but in reverse. Instead of turning images into sound, I guided Claude to use code to paint pictures with audio frequencies while composing the audio at the same time. Faces, birds and spirals. No DAW, MIDI or samples.
#CreativeCoding #ClaudeAI
Raymond Weitekamp retweeted
First a CPU, then a GPU. A computer isn’t complete without storage.
Meet qwenFS: Qwen3.5-2B with a filesystem compiled directly into its weights. Nothing runs outside the model. It adds 8 layers in front of Qwen’s own 24, using attention to hold the files and Gated DeltaNet state to hold the tables.
Every save and read runs in the model’s own forward pass. A write stores its text in the copy layer’s attention log, keyed by file, version and offset. A read uses those same keys, so attention copies the text back exactly token by token.
Meet qwenGPU: Qwen3.5-2B with a SIMT GPU inside its weights, 45 compiled layers in front of Qwen’s own, using its Gated DeltaNet layers as registers and memory.
It runs real CUDA in its forward passes. The host compiles the kernel to PTX and loads it, then the model writes every step of the launch itself: every register, branch and memory write.
The GPU has 4-lane warps, a 4-stage pipeline, 15 registers per lane and 126 words of global memory, all held in DeltaNet state. Qwen generates each launch token by token, each record running one machine block for one warp.
Raymond Weitekamp retweeted
There's a buzzy new paper about GPT-6 Astra being able to directly interpret model weights (prompts full of floating point numbers) - I tried it with Qwen 3.8 27B on my laptop and found that, after 45+ minutes, that model could achieve the same thing simonwillison.net/2026/Nov/3…
Raymond Weitekamp retweeted
i've reproduced the results from the MIT blog post today on weight-space programming. but one additional result i have to add: this is a new, step-change emerging capability that's appeared in just the latest set of models.
Raymond Weitekamp retweeted
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
this is essentially the "beam search without the beam" example i did on recursivedecisionmodels.com aka "smart binary search", with each pin deciding only left or right
wow i'm glad i listened to the end, i made it in with two minutes to go...
my original joke about @DSPyOSS was "shit, now i have to know what i actually want"
glad to see the joke spreading, thanks @dbreunig !
Context Engineering with DSPy xitter.gq/i/broadcasts/1qKVmydom…
rt-ing this because it is so cool and will be the death of my productivity
I used Divyansh's approach to build an entire Nintendo into ~1GB of Qwen3.5 weights.
Qwenputer, meet Nesformer: github.com/dbreunig/nesforme…
well i guess us omarchy folks are ngmi
中国社区逆向拆解了Claude的封号机制。以下是他们的发现
Claude使用积分制。注册时账号获得初始分数，之后根据行为加分或扣分。低于阈值直接封号。回分很难
影响分数的因素:
注册
> 干净IP是最重要的因素。不一定要家宽，但必须干净
> Gmail已经不是最优选，有权重更高的邮箱服务
> 注册时直接弹KYC的号可以扔了
> 用手机App注册会加分
支付
> 刚注册就订阅会扣分
> 直接上Max会大幅扣分。最好等两天先订Pro
> 支付时弹KYC说明账号在考察期，但不一定会封
> 国外实体银行卡加分。Google Play也行。虚拟卡扣分
使用
> 不要一次把限额用完。宁可两个号轮换也不要一个号撸到底
> Desktop客户端有少量加分。尽量少用网页端
> 手机端也有少量加分
> 开启记忆功能加分。记忆文件越多分越高
IP和设备
> 稳定性是关键。不要跳IP，尽量一直用同一个
> 代理必须全局模式并开启TUN
> 多设备登录大幅扣分。理想状态是一台电脑一个客户端
> 频繁换IP、频繁登录、高并发都会大幅扣分
账号年龄
> 老号有巨大的容错空间。2023年的号明显更高
> 分数随时间增长，但很慢
其他
> Windows和macOS没区别。Linux会扣分
> 时区和语言基本不影响。不会因为北京时间就封号
> 跟Claude说自己是中国人没问题，记忆不会触发封号
> 大量使用Fable会扣分
> 升级Max大概率进入观察期，除非是高权重老号
观察期和封号
> 接近阈值时账号可能进入观察期。典型特征是掉登录
> 观察期继续扣分就会封号
> 高分号可以扛过多次观察期
关注我，拆解你每天在用的服务背后到底怎么运作的
Raymond Weitekamp retweeted
Introducing 𝙻𝙵𝙼𝟹: our new generation of Liquid Foundation Models. With a new tri-hybrid architecture and new multimodal encoders, the LFM3 family is the ideal model suite for specialized intelligence.
> 6 model instances, across text, audio, vision, and omni modalities, generative and decision models.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱: text, 256k context window.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱: vision-language model with video-input support for spatial understanding
> 𝙻𝙵𝙼𝟹-𝙰𝚞𝚍𝚒𝚘-𝟸𝙱: tool calls, voice chat, reasoning in 18 languages.
> 𝙻𝙵𝙼𝟹-𝙾𝚖𝚗𝚒-𝟸𝙱: an omni model combines vision, audio, and language.
> 𝚍𝟷-𝟸𝙱 and 𝚍𝟷-𝙾𝚖𝚗𝚒-𝟸𝙱: decision models with typed outputs and calibrated probabilities in a single pass. Omni adds audio input.
> fine-tune on 1 GPU.
> a new tri-hybrid architecture
> architecture optimized for 14 processor classes across GPUs, CPUs, and NPUs.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱 is a beast. It performs competitively with Gemma 4 12B (6x larger), and Gemini 3.5 Flash Lite, and GPT-5.6 Luna, with 1.9B parameters.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱 is #1 on the Pareto frontier of device intelligence 𝚙𝚒𝚙𝚎𝚝𝚝𝚎.𝚕𝚒𝚚𝚞𝚒𝚍.𝚊𝚒
Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.
1/ 🧵
Raymond Weitekamp retweeted
USearch v3 is ready for a design-partner preview!
3 years ago, I released USearch v2. It became my first widely deployed piece of software.
It grew to over 4’000 GitHub stars, a number unimaginable to me back then. Together with my complementary projects, that’s 17’000 stars today. The open-source infrastructure behemoths shipping these libraries by default, like ClickHouse & ScyllaDB, collectively account for over 550’000 stars. Almost 800’000 including opt-in integrations 🤯
Thank you to everyone who shipped, starred, or argued with this code! I know I can be slow to respond, and I appreciate your patience.
With help from AI agents, compute from neocloud partners and a lot of user feedback, USearch v3 finally has a release candidate. Only 1’000 days of work later 🫠
USearch v3 stays open-source, Apache 2.0, free for everyone, from open-source giants to neolabs, hyperscalers, government agencies, & even mobile publishers!
Still, I’ve seen many times how paying for an integrated software experience can work better and cost less than getting the software for free — and be in everyone’s best interest. So I want to take a page from ClickHouse and Databricks, two organizations that reshaped commercial analytics infrastructure over the past decade by building businesses around open-source foundations.
I want to do the same for Search: keep the infrastructure open and productize the experience around it.
Before putting a major breaking release on “main” branch and launching an integrated offering around it, I want to make the transition as smooth as possible for my users.
So I’m inviting a few design partners to join me in final benchmarking and help shape the UX and C ABI before the public release. Your workloads should have a say in what ships.
I have no doubt USearch v3 will be a “scale unlocker” for many organizations reliant on Search. With a little creativity and collaboration, it can become a “capability unlocker” as well.
If you’re working at a frontier intelligence lab, in autonomous driving, or in life sciences, and managing 100+ TB of data, DM me with a brief description of your workload. I’d love to collaborate!
ALT github.com/unum-cloud github.com/ashvardanian
Raymond Weitekamp retweeted
in a twist of irony their built in packet logger has proven invaluable in capturing plaintext data meant for samsungs servers, fully mapping out telemetry it sends.
Raymond Weitekamp retweeted
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concerning-b…
Raymond Weitekamp retweeted
We finally have an explanation for “Facebook is listening to me through my phone”. They’d never do that (and it’s not really possible).
Turns out literally everyone else was — and then selling that to ad data brokers.
WTF, $3000 Samsung TV has an always-on mic, automatic content recognition, geo tracking, network mapping, and ships your data to Samsung + 400 partners.
Bro Rooted a brand-new 2025 Samsung TV after noticing absurdly targeted ads.
Firmware was locked down, so zero-days it was.
Got a root shell.
then found: always-on voice service, ACR, geo, full telemetry pipeline, and four separate DRM systems.
This isn’t a bug, It’s the product bro.
