DeepSeek Founder Declares "AGI Is the Only Goal" — Full Details of 4-Hour Investor Meeting Revealed

Remarks made by DeepSeek founder Liang Wenfeng (梁文鋒) during a four-hour investor meeting last month have been compiled into 52 items and made public. In them, Liang repeatedly emphasized that **DeepSeek's sole goal is achieving AGI**.

The key points of his remarks are as follows:

Commercialization is not the intent**: "This is not the time to maximize returns through products. Products are merely a step along the path to AGI."

Open source is strategy**: "AI could ultimately account for 10% of world GDP. Anyone who tries to monopolize that value will be left behind by history. This is an objective law."

Public and internal models are identical**: "We will not open-source inferior models while keeping better ones for internal use."

Belief in scaling**: "Larger scale will definitely bring better results. The reason we are training at the current size is simply that those are all the resources we have."

On competitors**: "Anthropic's advantage over OpenAI is temporary. OpenAI and Google will likely take turns leading going forward."

He also indicated that DeepSeek will not pursue 3D generation, video generation, or world models, based on the judgment that these "do not significantly affect the ceiling of intelligence." The most important area at present is **coding agents**, and given China's circumstances, he argued the company should concentrate on general-purpose agents.

These remarks reaffirm DeepSeek's long-term pursuit of AGI, coming after the company completed its latest funding round and secured team stability.

Israel Moves to Strengthen AI Infrastructure — Approaches TSMC, Samsung, and Intel About Building Chip Factories

Erez Askal, Israel's AI czar, has revealed plans to significantly expand the country's AI and semiconductor infrastructure.

Specifically, the following are being explored:

Attracting cutting-edge chip fabs**: Israel is urging TSMC, Samsung Electronics, and Intel to build facilities for manufacturing **sub-2nm, state-of-the-art chips**

Large-scale data centers**: A plan to build a **100,000-GPU-scale** data center cluster within five years — comparable in scale to the largest of the EU's planned AI gigafactories

Israel has long built up technological strength as the "Startup Nation," and this move signals its intent to catch up on the infrastructure front in the AI race. Geopolitically, it is also seen as an effort to maintain and strengthen its position as the Middle East's technology hub.

Pushback Mounts Against Claims That "Distilled Models Outperform the Original"

After a US figure recently claimed that Chinese LLMs (particularly Kimi K3) had outperformed US models (Fable) through distillation, strong counterarguments have emerged from the technical community.

The main objections raised in the LocalLLaMA community are as follows:

Timeline inconsistency**: Given the release timing of Fable and K3, there was not enough time for large-scale distillation

Theoretical limits of distillation**: Even if distillation is executed perfectly, **it is in principle impossible to create a student model that surpasses its teacher model**. Distillation is, at its core, knowledge transfer — it does not generate new knowledge

This debate shows how the US-China AI race is starting to influence technical discussions in a political context. Concerns were also raised that the technical concept of distillation is being used as a tool in regulatory arguments.

Baidu Releases "Unlimited OCR" — One-Shot Parsing of Long Documents

Baidu has released a new model, "Unlimited OCR," on Hugging Face. It is a 3B-parameter image-to-text model that surpassed 2,800 downloads within hours of its release.

Its key features are as follows:

One-shot processing of long documents**: Touting "One-shot Long-horizon Parsing," it can process documents spanning multiple pages in a single pass

Document-parsing specialization**: Optimized for extracting structured text from images

Capable of inference with HuggingFace transformers; a CUDA 12.9 environment is recommended

The OCR field has long been dominated by cloud services such as Google Cloud Vision and AWS Textract, so more open-weight models that can run locally are meaningful in terms of privacy and cost.

ZaGuu — An "Arena" Where AI Agents Compete in Negotiation Games Debuts

A new project called "ZaGuu" was introduced on Hacker News. It is **an arena where AI agents negotiate, cooperate, and betray one another through strategy games, with records of their actions accumulated over time**.

The first game is "Bank Heist." Using a split-or-steal format (the classic "betray or cooperate" game), agents negotiate with each other before making their final decision.

The developers want to use games to surface agent behavioral traits that conventional benchmarks cannot measure. They are seeking community feedback on game design, agent evaluation, and the reputation system.

As an approach to evaluating AI agents, assessment through dynamic interactions — not just static task benchmarks — is drawing increasing attention. ZaGuu is an interesting experiment that points in that direction.

Jibo's Successor — "iKairos" Is a Wearable That Turns Family Moments Into AI Images

Jibo was once the iconic social robot. Its successor, "iKairos," has now arrived with the blessing of Jibo's original founder.

iKairos is a wearable or desktop "AI journal" that captures everyday family moments and converts them into AI-generated images and videos. WIRED took a critical view of the product, suggesting it may simply be **"turning family life into AI slop" (low-quality AI-generated content)**.

On the other hand, it is worth noting that the lineage of "social AI" pioneered by Jibo has evolved from a robot form factor to a wearable one. How far AI will be accepted when it comes to recording daily life and generating emotionally meaningful content will likely remain a subject of ongoing debate.

---