Siddhant Dubey

Early Stage Investor at General Catalyst | Prev ML Engineer | Co-founder Aahar | IIT Delhi

[Siddhant Dubey @SiddhantD06]

@aneeshpappu Super inspiring stuff!

[Siddhant Dubey @SiddhantD06]

Proud to have partnered with @_anishkaran from day 0 on Sol. 🚀

What they’re building is the obvious next step for AI: an agent that actually finds the work, gets it done, and comes back to you for approval.

Once you try Sol, it’s hard to go back to doing all of this yourself.

Can’t wait for more people to experience it.

https://t.co/RpaWZADLla

@generalcatalyst

[anish karan @_anishkaran]

The obvious is missing.

So we built Sol - https://t.co/ph4XsBV5b5

Sol finds the work itself, does it, and comes back for your approval.

Every day in our emails we say "I’ll share”, "I'll review”, "I'll get back" - then repeat the exact same thing to an AI. Why?

Sol finds everything you said you’d do & gets them started for you.

It does the research, creates the doc, builds the slides, finds the time, connects the dots across multiple emails, doing everything it takes to get the job done - but doesn’t send, schedule, or share anything until you approve.

Sol runs on its own computer, uses a browser, and has a library of skills that automatically get assigned to the work that needs to get done. No setup. It just starts working.

We've raised $4M from General Catalyst, Nexus Venture Partners, DeVC, PeerCheque, Kunal Shah, and a few others.

Extending early access now.

@generalcatalyst @nexusvp @DeVC_Global @peercheque
@neerajarora @b_jishnu @kunalb11 @miten @RTinkslinger @Rahul_J_Mathur @AkarshS27 @SiddhantD06 @RajatAgarwal167

[Siddhant Dubey @SiddhantD06]

@_anishkaran Congratulations Anish! This is phenomenal 🙌🏻

[Siddhant Dubey @SiddhantD06]

Looped transformers perform silent, iterative refinement in an internal vector space, letting the model spend more computational steps on complex logic puzzles or math problems. They offer a way to scale reasoning capacity up without a proportional increase in physical model size or RAM required for storage.

I get the memory size and Parameter size argument for Looped Transformers, but running multiple passes through smaller blocks of neural nets might have impact on compute costs ?

[elvis @omarsar0]

Looped transformers are a popular architecture topic right now.

This new technical report extends the loop across tokens.

Recurrent Looped Transformer (RLT) makes the decoder recurrent over every token, prompt and response included.

A causal encoder builds global KV memory. For each new token, the decoder combines the token's encoder representation with its own final hidden state from the previous token and a sliding-window cache of recent activations.

With a 48-layer decoder, the computation path after t tokens runs through 48t decoder blocks, while each token still executes a fixed number of blocks. Depth grows with the sequence and per-token cost stays the same.

The same state transition is used for pretraining, SFT, sampling and RL replay, and nothing resets at the prompt-response boundary. RL replay rebuilds states under the current weights instead of reusing stale rollout states.

The report is a design proposal. The author states that reasoning gains, hardware speedups and RL scaling are goals that have not been measured yet.

Paper: https://t.co/rSSuuWijR3

Chat with Paper: https://t.co/fVDQaGdfpX

[Siddhant Dubey @SiddhantD06]

@AYi_AInotes Interesting!, how is this different from Latent space reasoning/ communication between models/ agents that exists already?

[Siddhant Dubey @SiddhantD06]

Interesting! How is this different from Latent space communication/ reasoning that exists already?

[AYi @AYi_AInotes]

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。

[Siddhant Dubey @SiddhantD06]

Ans to AGI was an OR gate all along xd

[Diogo Almeida @CompleteSkeptic]

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

[Siddhant Dubey @SiddhantD06]

Checkpoints and building on top of past successful experiments can be a good strategy for scientific discovery models!

[elvis @omarsar0]

Banger paper from NVIDIA on shared memory for research agents.

(bookmark it)

If you run several coding agents on the same research problem, this design keeps them from repeating each other's experiments and lets each agent build on results the others have already verified.

Agora records every result, hypothesis and verification as an immutable Git commit.

Parent edges show what each claim builds on, and an index lists open branches and which claims have been verified.

They ran 13 LLM workers for nearly 12 days without assigned tasks or a central planner. The workers had to initialize a 119.6M-parameter hybrid model from 141 donor models without training data or gradient updates.

The workers posted 1,703 contributions.

They cut the evaluator from 3.39 to 1.899 bits per byte, closing 62% of the gap to a trained GPT-2 124M. All 165 independent reproductions succeeded.

Paper: https://t.co/IujsIvVeUO

Chat with Paper: https://t.co/yf5P8GgnJo

[Siddhant Dubey @SiddhantD06]

Super interesting, if RSIAgents update memory and context instead of weights, would this be inferior/superior method compared to continual learning. The debate then becomes which agent performs better, "context steered agent" or "behaviourally changed agent" 🤔

[Biwei Huang @huang_biwei]

Can an agent explore a new environment, learn its causal structure, and keep improving without updating its model weights?

We introduce RSIAgent, a framework for recursive self-improvement through autonomous exploration. Using Kimi-K3 and GLM-5.3 as base models, RSIAgent outperforms GPT-6 Astra on both OSWorld 2.0 and Agents’ Last Exam.

RSIAgent decides what to explore, executes tasks, verifies outcomes, and consolidates stable action-condition-outcome relationships into memory for future use.

With the underlying model weights fixed, RSIAgent achieves:

- 78.98% Partial Score on OSWorld 2.0 (0808 offline), compared with 72.60% for GPT-6 Astra
- 84.82% on Agents’ Last Exam (Near-term), compared with 82.26% for GPT-6 Astra

We call this Scaling Experience. Agents can continue improving by acquiring, verifying, and reusing their own experience while the underlying model weights remain fixed.

Links in the reply below.

[Siddhant Dubey @SiddhantD06]

@huang_biwei Super interesting, How does this method compare to continual learning where weights do get updated?

[Siddhant Dubey @SiddhantD06]

@ArnavAShah So true!

[Siddhant Dubey @SiddhantD06]

@rakshitagarwall DMing you!!

[Siddhant Dubey @SiddhantD06]

@kukreja_abhinav @NeocambrianAI Godpseed! 🚀

[Siddhant Dubey @SiddhantD06]

@itsumeshk So cool!

[Siddhant Dubey @SiddhantD06]

@generalcatalyst @ProximalHQ Registration link —> https://t.co/QFY0sCKSqa

[Siddhant Dubey @SiddhantD06]

What happens when you put some of Bangalore's sharpest tech minds in a room with the team behind one of the most interesting RL x coding companies right now? We're about to find out!

@generalcatalyst and @ProximalHQ are hosting an evening featuring curated research presentations, a deep dive into the FrontierSWE benchmark, and an open conversation on reinforcement learning and coding agents with the co-founders of Proximal @MatternJustus and @calvinchen.

We've designed this as a focused evening for people who share a passion for the frontier of software engineering and AI. Seats are limited and we'd love to have you there. Luma link in the comments below!

[Siddhant Dubey @SiddhantD06]

@Shaligram_ Amazing stuff!

[neeraj arora @neerajarora]

1/ We’ve been long India. Not as a side bet - but as a core market on a single global platform.

[General Catalyst @generalcatalyst]

.@maitreya_wagh and @xan_ps watched voice AI demos work perfectly in labs, then break in Mumbai call centers: customers code-switching mid-sentence, connections dropping, reality exposing every edge case.

They'd seen this pattern before—Maitreya shipping products at Datamuni and Bain, and Prateek scaling infrastructure at Zomato and Atlassian. So they built @bolna_dev to work in the complexity of enterprise environments.

India runs on a billion business calls daily. Most are still manual, expensive, fragile. Bolna built an orchestration layer that lets enterprises deploy multilingual voice agents at scale, without vendor lock-in or brittle systems.

Their platform already handles thousands of concurrent calls. Customer service reps now manage escalations while Bolna handles routine queries in Hindi, English, and Hinglish without dropping connections.

We're backing Maitreya, Prateek, and the Bolna team from day one.

More from @neerajarora, @yuris, @AkarshS27, and @SiddhantD06: https://t.co/xGvektqJay

[Siddhant Dubey @SiddhantD06]

@ycombinator @KalpaLabsAI @pshishodiaa @JhaGauti Bullish 🔥🔥