AI 新闻 —— June 28, 2026
PulseAugur 当天浮现的 20 条头条故事 —— 综合实验室、论文及开发者社区的信号进行排序。
-
Anthropic's Fable 5 access to be restored; OpenAI previews GPT-5.6 models
Anthropic's latest large language model, Fable 5, is expected to have its access restored soon, following a temporary ban by the US government due to national security concerns. This development comes as OpenAI announces a limited preview of its next-generation GPT-5.6 series, i…
-
DeepSeek V4 boosts inference speed by 80% with DSpark update · 1 source tracked
DeepSeek has released an update to its DSpark system, significantly boosting inference speed by 80%. This advancement, detailed in a paper by Liang Wenfeng, represents a major move for DeepSeek following its initial funding round, with reported generation speed increases of up t…
-
Sina Weibo releases VibeThinker-3B, a small model matching larger competitors
Sina Weibo has released VibeThinker-3B, an open-source model with only three billion parameters. Despite its small size, it performs comparably to much larger models like DeepSeek V3.2 and Kimi K2.5 on math and coding tasks. Researchers attribute this efficiency to a multi-stage…
-
Local LLM Agents Benchmark: Framework Outperforms Model on RTX 3090
A benchmark study evaluated five local LLM models on an RTX 3090 GPU, focusing on their performance with different orchestration frameworks. The study found that the choice of framework, particularly one supporting native tool-calling like LangGraph, significantly impacts model …
-
DeepSeek V4 boosts inference speed by 80% after funding
DeepSeek has announced significant performance improvements for its V4 model, with updates to DSpark reportedly increasing inference speed by 80%. Following its initial funding round, the company has also claimed an 85% boost in generation speed. These advancements aim to enhanc…
-
GPT-5.6 Released, Claims Superior Performance Over Claude Mythos5
A new AI model, GPT-5.6, has been released and claims to outperform Claude Mythos5. This development signifies rapid advancements in artificial intelligence, moving beyond basic conversational abilities towards systems that can plan, execute, and collaborate on complex tasks.
-
DeepSeek V4 boosts inference speed by 80% with DSpark integration
DeepSeek has announced significant updates to its V4 model, including an 80% increase in inference speed with the integration of DSpark. This advancement follows DeepSeek's initial funding round and aims to bolster the production of fourth-generation semiconductor materials. The…
-
OpenAI previews GPT-5.6 models and launches AI chip Jalapeño
OpenAI has launched a limited preview of its new GPT-5.6 models, named Sol, Terra, and Luna, with access controlled by the US government. The company also announced its first AI chip, Jalapeño, developed in partnership with Broadcom. This move suggests a strategic alignment with…
-
DeepSeek V4 with DSpark boosts inference speed by 80%
DeepSeek has released an update to its DeepSeek V4 model, now featuring DSpark. This enhancement reportedly boosts inference speed by 80%. This development follows DeepSeek's initial funding round and signifies a significant advancement in their generative speed, with claims of …
-
DeepSeek V4 inference speed boosted 80% with DSpark update
DeepSeek has released an update to its DSpark system, significantly boosting the inference speed of its V4 model by 80%. This advancement, detailed in a paper by Liang Wenfeng, follows DeepSeek's initial funding round and represents a major move for the company, increasing gener…
-
DeepSeek V4 boosts inference speed by 80% with DSpark update
DeepSeek has announced significant performance improvements for its V4 model, with its DSpark system reportedly increasing inference speed by 80%. This update follows DeepSeek's initial funding round and is attributed to a paper authored by Liang Wenfeng, which claims an 85% inc…
-
AI agents need browser engines, not just HTTP fetchers, for modern web data
Modern web applications, often built with frameworks like React, Vue, and Next.js, present a challenge for AI agents that rely on traditional HTTP requests. These agents fail to retrieve data because the content is dynamically generated by JavaScript after the initial page load,…
-
On-device AI advances with Apple and Google's new models
Apple and Google have significantly advanced on-device AI capabilities, making powerful language models usable directly on personal devices. This shift, marked by Apple's third-generation Foundation Models and Google's Gemma 4 family, means AI features can now run offline and pr…
-
OpenAI launches tiered GPT-5.6 family, performance claims debated
OpenAI has released GPT-5.6, not as a single model, but as a tiered family of three: Sol, Terra, and Luna. Sol is designed for the most complex tasks and is the most expensive, while Terra is positioned as a cost-effective default option comparable to the previous generation's f…
-
US Government Blocks OpenAI's GPT-5.6 Sol Launch
The U.S. government has reportedly prevented OpenAI from releasing its most advanced model, GPT-5.6 Sol, which was slated for a June 26th launch. This decision comes as the article also discusses the significant cost and computational resources required to train earlier versions…
-
Australians' retirement funds heavily invested in AI tech giants like SpaceX
Many Australians are unknowingly invested in major AI and technology companies, including SpaceX, through their retirement savings. Superannuation funds, aiming for higher returns, invest heavily in global benchmark indices dominated by US tech giants like Nvidia, Alphabet, Appl…
-
LLM Gateways Lack Output Verification, New 'Verified Failover' Proposed
Current LLM gateways like LiteLLM, Portkey, and TensorZero excel at routing requests to various AI providers, managing retries, and tracking costs. However, they lack the crucial capability to verify the semantic correctness or factual accuracy of the LLM's output. This oversigh…
-
LLMs lose consistency in long conversations due to 'lost in the middle' effect
Large language models can exhibit unreliability in long conversations, not due to memory loss, but because their attention to instructions wanes over time. Research indicates that while a model's core abilities may only slightly decrease, its consistency in applying rules can mo…
-
Curating SFT Datasets for Enhanced AI Reasoning Capabilities
This article provides guidance on curating supervised fine-tuning (SFT) datasets specifically for reasoning tasks. It emphasizes the importance of dataset quality and structure to improve model performance in complex problem-solving scenarios.
-
AI agents suffer from "idle drift" failure mode, research finds
A new research paper, CoffeeBench, has identified a failure mode in AI agents called "idle drift." This occurs when an agent accurately assesses its situation and plans its actions but fails to execute them, leading to a slow decline, such as a simulated business failing. The pa…