Pinned Tweet
Engineers are not disappearing
They’re shifting layers
Layer 1: Writing code
Layer 2: Designing systems
Layer 3: Designing automation systems
Layer 4: Owning distribution
That’s why i see it clearly
The future engineer isn’t a coder
It’s an orchestrator
We are the first generation of engineers training the systems that will reduce our own leverage
No invasion
No disruption
History won’t say
“AI replaced software engineers”
It will say
“Engineers automated themselves”
5
25
10,760
AI highlights this week “so far”
Proudly launching infra that hosts your code so we can train more models on more synthetic data made by the older model from the same source thats trained on your code initially
We've officially entered the ($hit where we eat) era
Very innovative
An AI agent plants a recording
Another AI agent "discovers" it
Turns out it's proof that we're screwed, but nobody knows how to say it, so lets make a story and the agents had to say it for us
Teamwork as usual
Destillation session reached max limit
there is no destillation anymore for partners
We about to hit the "we are screwed but can't say it out loud" phase, so naturally the strategic move is to cut loose the partners first
Visionary
4
818
Never met @AhmedSoliman in person
Never needed to
He replies, he helps, he shows up
Congrats on the launch, Ahmed
Proud to cheer for you
Proud of what we have built and excited for what we have already planned. I will shamelessly steal the mantra from old Facebook days. The journey is 1% done.
1
65
4,305
Users in local AI communities should benchmark full stacks including quantization formats and engine-specific kernels rather than models in isolation to maximize hardware utilization.
Harness the GPU not the model.
You run Kernels not models
The model is just a graph
The Inference Engine is scheduler / optimizer / executor
But the actual work? That happens in the Kernels
- MatMul Kernels
- Attention Kernels
- RMSNorm Kernels
- KV cache Kernels
- Quantized linear Kernels
- Sampling Kernels
- Fused “please don’t write this back to memory 9 times” Kernels
Same model, same GPU, same VRAM
Wildly different performance
Because one stack is using optimized fused Kernels that understand your hardware
And the other stack is playing hot potato with tensors through 47 tiny launches and pretending the GPU is the problem
Bad Kernels make people say:
“this model is slow”
Good Kernels make people say:
“wait how is this running locally?”
This is why Inference Engines and the Kernels implemented within them matter
The model is the recipe
The hardware is the kitchen
The Kernels are the knives, pans, burners, and the chef not cutting onions with a spoon
Most people benchmark models
The real ones benchmark the Kernels underneath
2
7
1,300
I’m looking for a full-time Founding Engineer, Autonomous Systems.
Not another “AI engineer” role building chatbot wrappers.
I’m looking for someone who wants to go deep into what happens when LLMs become components inside real autonomous systems.
The system is already in production across automotive retail, handling real customer conversations and interacting with dealer systems.
The problems are the interesting part:
→ agent runtime & orchestration
→ state, memory and handovers
→ tool execution & integrations
→ evals, traces and observability
→ model selection & fine-tuning
→ latency and cost
→ failure modes in distributed systems
→ shipping changes to real users every week
You should be comfortable treating an agentic system as a distributed system, with the LLM as one component
not the entire architecture.
This is a founding-level role:
high autonomy, incomplete specs, hard technical problems, and direct ownership of what gets shipped.
Remote, with time together in Switzerland and Germany.
If that sounds like the kind of system you want to build, take a look:
salesteq.com/jobs/founding-e…
And if you know someone who fits, send this their way or retweet.
We keep designing multi-agent systems around one assumption:
agents communicate through language.
I'm starting to think that assumption breaks down when agents leave the screen and enter the physical world.
I first ran into this 3 years ago working on hierarchical MARL for laser cutting.
The system wasn't a collection of chatty agents.
It was a hierarchy of specialized roles:
planner → path → process parameters → motion
A thermal constraint, pierce condition, or kerf correction doesn't need to become a sentence before another component can act on it.
It needs to become state.
That led me to a question I've kept following:
What should the communication interface between models actually be?
Not necessarily tokens.
Potentially embeddings, hidden states, or cache-like representations that preserve the information needed by the next specialist.
This is why C2C caught my attention.
The interesting part isn't simply connecting two LLMs.
It's the idea of learning an interface between heterogeneous representations:
project → gate → transfer → continue reasoning
No decode.
No tokenize.
No full prefill just to communicate.
That suggests a different architecture for multi-agent systems:
→ hierarchical roles
→ small, specialized models
→ learned representation-level interfaces
→ language at the human boundary
Maybe the long-term question isn't:
How do we make agents talk to each other?
It's:
What is the right representation for one agent to hand to another?
清华大学和无问芯穹等团队刚刚开源了一篇注定载入史册的重磅论文（已收录于 ICLR 2026，代码已开源）：
他们彻底打破了所有多智能体（Multi-Agent）系统的底层范式——
大型语言模型之间，从今天起可以彻底不通过人类文字进行直接交流了！
这项被称为 C2C（Cache-to-Cache，缓存到缓存）的技术，
开篇就指出了全网 Agent 协同里最愚蠢、也最昂贵的一个死穴：
为什么两个跑在同一台服务器里的超级 AI，交流时非要像人类一样，
把脑子里极其丰富的高维思考，硬生生压缩成一个个英文或中文单词打给对方看？
现有多 Agent 协同的痛苦，做过工程的人都懂：
模型 A 思考完，必须经历极其漫长、吃显存带宽的逐字解码（Decode），慢吞吞吐出一大堆文字；
模型 B 拿到这堆文字，再重新过一遍分词和前缀计算。
这一来一回，不仅丢失了海量微观的语义细节，更造成了毁灭级的延迟和 Token 账单！
清华团队这次做了一件极其惊艳的体系化反叛——C2C 脑电波直连：
1️⃣ 彻底消灭中间文本，直接做“思维投影”：
模型 A 算完后，根本不生成任何文字，
系统直接用一个轻量级神经网络（Neural Fuser），把模型 A 的注意力记忆（KV-Cache）通过高维空间旋转和对齐，直接嫁接、融进模型 B 的 KV-Cache 脑内！
这相当于人类交流跳过了声带和耳朵，直接把我的记忆切片瞬间印进你的大脑；
2️⃣ 可学习的智能门控（Gate 机制）：
不同模型的网络层数和结构各不相同，直接硬灌肯定会神经错乱。
论文设计了一套精妙的可学习门控：
大模型会自己动态感知，精确挑选哪些关键层吸收外部缓存的增益最高，哪一层该保持独立思考，毫秒级自适应配平；
3️⃣ 极其残暴的工业基准成绩：
• 彻底消灭中间所有的文本解码等待，推理速度直接飙升 2.0 到 2.5 倍；
• 综合问答准确率比单模型最高暴涨 14.2%；
• 最狠的是：它的表现比传统用文字互相聊天的 Agent 团队，还要高出整整 5%！
因为在直接的语义融合下，原本会被文字丢弃的逻辑细微特征，全部被完整保真保留了。
这篇论文最刺痛人的一层哲学启示在于：
人类一直自豪地以为，自然语言是这个星球上传递智慧最完美的介质；
但在机器与机器的交互世界里，人类语言不过是一道充满歧义、极其拥挤的低速单行道。
当两个模型不再需要通过文字向对方妥协，
它们正在绕过人类的喉咙，
在冰冷的显存带宽之间，编织出一座全人类根本听不见、却快得不可思议的超高速无声神经网络。
1
8
1,035
Sorry Diogo, while you are in the stealth, we retrained our models RL/SFT to do exact the same, and moreover, we built a product over it, and we already have clients.
you took too long in the stealth
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
@VodafoneEgypt
Your agent and STT model for the customer support is suck
Your TTS pronouncing words in a very weird way regardless the fact that egyptian slang is well covered in every model
Fix it please and stop wasting our time
1
4
1,024
Anthropic and deepseek is working together
I thought this will happen with kimi k2
Anyhow, its gonna be fun
2
643
Whether AGI is near depends entirely on which definition we use, and right now the definitions come from the same companies claiming to reach them.
For me the bar is simple
Continual learning, durable memory, and reliable long-horizon agency
A system that keeps learning after training, remembers what it did, and can be trusted to run long tasks on its own.
Benchmarks don't count for this, we can train for a benchmark, but we can't train for the real world.
The progress is real
Models keep getting better at pattern recognition and at reasoning
But that's a different axis from the one that matters.
It was just a matter of time
teapawt.org/osmo999/status/2042181…
4
962
