DeepSeek Rejects Alibaba Partnership, Prioritizing Corporate Independence

It has emerged that DeepSeek walked away from investment talks with Alibaba. When DeepSeek announced a large-scale fundraising plan in April, both Tencent and Alibaba — China's two biggest tech giants — expressed interest, but no agreement could be reached with Alibaba on the terms.

Alibaba pushed hard for integration into its own AI ecosystem, while DeepSeek prioritized keeping restrictive clauses to a minimum and preserving its position as an independent model company. DeepSeek reportedly has no shortage of other interested investors, leaving it plenty of options.

The move throws into sharp relief the "big tech umbrella vs. independent" divide in China's AI industry. On top of its technical prominence, DeepSeek has now made its independent path unmistakably clear as a matter of corporate strategy.

Yoshua Bengio Declares "I've Solved AI Alignment"

Yoshua Bengio, one of the pioneers of deep learning, has made an announcement titled "I've Solved AI Alignment." The details have been published as a YouTube video.

Bengio has spent years warning about AI safety and alignment, and the declaration has drawn a strong reaction. That said, the specific content and methods behind this "solution" will have to wait for the detailed published materials. For now, while some in the industry remain skeptical, others argue that given Bengio's track record, the claim cannot be dismissed lightly.

llama.cpp + MTP Runs Qwen3.6 35B A3B at 80 tok/s on 12GB VRAM

A report gaining attention in Reddit's LocalLLaMA community says that by using a PR that adds Multi-Token Prediction (MTP) support to llama.cpp, the Qwen3.6 35B A3B model achieved generation speeds of more than 80 tok/s on a 12GB VRAM setup.

On a configuration using an RTX 4070 Super, the draft acceptance rate topped 80%. It also supports 128K context, making the prospect of running large models at high speed on consumer GPUs suddenly feel within reach.

MTP is a technique that predicts multiple tokens ahead simultaneously rather than just the next one, contributing significantly to faster generation. If integration into llama.cpp moves forward, it is expected to benefit many local LLM users.

Typewhisper: A Privacy-First Local Speech Recognition AI

An open-source local speech recognition tool called "Typewhisper" is making waves on Hacker News. It performs AI-driven speech-to-text on-device, featuring a hybrid design that safeguards privacy while still allowing cloud use when needed.

Because audio data is never sent to external servers, it can be used with confidence even when handling sensitive information. It is available for Mac and looks well suited to everything from everyday transcription to specialized applications.

AI CAD Tools Struggle with Basic Mechanical Parts — CADBench

A benchmark project called "CADBench" is drawing attention on Hacker News. A developer put the major AI CAD tools through their paces, and every one of them failed at designing basic mechanical parts.

While AI-driven code generation and text processing are advancing rapidly, the results highlight that in CAD — a field that demands adherence to physical constraints and precise dimensional control — human engineering judgment remains indispensable. It is a fascinating attempt to clearly delineate the boundary between AI's strengths and its weak spots.

AI Is Transforming Cement Manufacturing — Washington Post Reports

The Washington Post has reported on the transformation AI is bringing to cement manufacturing. The cement industry is a massive sector accounting for roughly 8% of global CO2 emissions, and AI-driven process optimization is expected to deliver both improved efficiency and a reduced environmental footprint.

AI adoption in manufacturing is expanding beyond software into physical industrial processes as well.