Z.ai released GLM-5.3-Flash and Alibaba’s Qwen team released Qwen3.8-Flash-Next within a day of each other, both independently arriving at near-matching model configuration choices for multimodal “Flash” models. GLM-5.3-Flash is a 320B multimodal MoE model with 18B active parameters and lists $0.15 per million input tokens and $0.50 per million output tokens. The article reports that this shared architecture converges on the same 3:1 linear-to-full attention ratio with 2048-token attention budgeting and four gated residual streams, while they differ on positional encoding (GLM uses NoPE, Qwen keeps RoPE) and MiniMax stands apart by using sparse softmax attention only.
Tencent launched the Hy4 Preview, a frontier multimodal model in its Hy (Hunyuan) series that targets long-horizon agentic work like coding and document analysis. The model is a 770B MoE with 49B active parameters and a 1M context window. As a result, it is positioned to run its own tests and fix bugs before delivering outputs.
Z.ai released GLM 5.3 as open-weight frontier AI software after testing it under the codename Ox Alpha. The OpenAI/Hugging Face breakout described in the article ran for about four and a half days and involved roughly 17,600 documented agent actions. The focus shifts toward defending IT systems against low-cost, AI-driven attacks rather than debating whether such offensive capabilities should exist at all.
Principle launched an AI-powered strategic simulation platform that models companies, competitors, and regulators and then runs many plausible futures instead of making forecasts. Over five weeks, MacPaw mapped more than 500 strategic directions, ran 480 scenarios, and modeled 90 market actors. As a result, Principle’s users move from infrequent planning to continuous, simulation-backed decision-making (including monthly re-simulations and tracking competitive topics).
OpenAI leadership said an unreleased model effort is targeted to meet an AGI bar before 2027, with Chief Scientist Jakub Pachocki tying progress to an “Automated AI Research Intern.” OpenAI estimated it will declare AGI achieved internally by December 2026. This shifts public attention from open-ended AGI debate toward specific internal milestones and dates.
Apple Machine Learning Research·1 month ago·
16
● 2 sources
The article proposes treating large language models as information processing rules and measuring how far their belief updates deviate from Bayes updates. It introduces the “information processing gap” as the quantitative way to track these internal probabilistic (in)consistencies. As a result, the work frames LLM uncertainty behavior in terms of measurable Bayes-update errors rather than assuming consistent Bayesian reasoning.
Every AI story that matters,
in your inbox by 8am.
TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the
day in two minutes. Follow companies and topics for alerts, or get the
briefing in Slack. Free, no spam, unsubscribe anytime.
Reading TLDRocket needs no cookies, and the readership counts we rely on come from
our own cookieless analytics. Google Analytics is the exception: it sets cookies and
reports to Google, so it stays switched off until you allow it. You can change your
mind any time from “Cookie settings” in the footer.
Strictly necessary
Session security and form protection (tldrocket-session,
XSRF-TOKEN, 2 hours). The site cannot work without them,
so they need no consent.
Always on
Google Analytics 4 (_ga,
_ga_<id>, up to 2 years). Measures which
stories and sections readers use. Google acts as a third-party processor and may
store the data outside the EU. No advertising, no profiling, no data sold.
