Research & Models · CN

The Download: DeepSeek’s latest AI breakthrough, and the race to build world models

DeepSeek V4 processes significantly longer context windows than the previous generation, closing a capability gap with frontier Western models.

DeepSeek released a preview of V4 on Friday, its long-awaited flagship model with extended prompt processing capacity, according to MIT Technology Review. The model architecture supports longer input sequences, enabling it to handle more complex tasks that require retaining and reasoning over larger document sets.

Research & Models · CN

1,6 Billionen Parameter & Open Source: DeepSeek V4 stellt den globalen KI-Markt auf den Kopf – Chinas nächster Angriff auf den KI-Weltmarkt - Xpert.Digital - Konrad Wolfenstein

DeepSeek V4's 1.6 trillion open parameters just forced every closed-model incumbent to recalculate their competitive moat.

Chinese AI lab DeepSeek released V4 as open source on 27 April 2026, claiming competitive performance at scale. The model's public weight release and parameter count let any builder replicate or fine-tune without licensing fees or API dependence.

Research & Models · CN

Intel warns China of severe server CPU shortage as AI demand surges

Intel warned Chinese customers of severe server CPU shortages as AI demand explodes past chip supply.

Intel issued a shortage alert for China in April 2026, signaling that AI workload demand has outstripped semiconductor manufacturing capacity. Datacenters training and running large models are consuming chips faster than fabs can deliver, creating a supply crunch that extends to enterprise customers.

Research & Models · CN

DeepSeek's new models are so efficient they'll run on a toaster ... by which we mean Huawei's NPUs

DeepSeek V4 runs inference on Huawei silicon at a fraction of the cost rivals demand, collapsing the economic moat around closed models.

DeepSeek released V4 in preview on 24 April, an open-weights LLM claiming competitive performance with frontier proprietary models while cutting inference costs dramatically and extending support for Huawei's Ascend accelerators. Architectural redesign handles longer context windows and reduces computational overhead, making the model viable on consumer-grade and domestic Chinese hardware.

Research & Models · UK

DeepMind’s David Silver just raised $1.1B to build an AI that learns without human data

David Silver, DeepMind's AlphaGo architect, raised $1.1 billion at a $5.1 billion valuation for Ineffable Intelligence, a startup aimed at AI that learns without human-labeled data.

Silver founded Ineffable Intelligence months ago and has already secured institutional backing; the lab claims to pursue reinforcement learning without supervised datasets. The company plans to use self-play and algorithmic optimization to replace human annotation as the data bottleneck.

Research & Models · UK

The Man Behind AlphaGo Thinks AI Is Taking the Wrong Path

David Silver argues that AI is pursuing the wrong path and that self-supervised learning without human data is the answer.

Silver, architect of AlphaGo, founded Ineffable Intelligence to build AI that scales without human annotation; the company raised $1.1 billion. Silver's thesis hinges on reinforcement learning and algorithmic self-improvement replacing supervised learning as the dominant training paradigm.