AI Briefing
TLDR AI news
The latest AI and developer news from TLDR AI, with short summaries.
Visit website →DeepSeek's New AI Model V4-Flash Has the Lowest Running Cost Among Popular Models… 105 Times Cheaper Than Claude Fable 5
DeepSeek V4-Flash was found to be 105 times cheaper to run than Claude Fable 5.
Google Narrows Feature Gap in Gemini Desktop App
Google is adding image and video generation and camera features, among others, to the Gemini desktop app.
When Geniuses Fail—Intellectual Arrogance at AI Labs
It points out that an AI expert's success does not guarantee expertise or achievement in other fields.
APEX-Accounting: An AI Productivity Benchmark for Accounting Work
APEX-Accounting evaluated whether AI can repeatedly produce accurate results in accounting tasks.
A New Era of AI Testing
Opus 5 implemented a Lord of the Rings scene in code in just 2 hours.
MSLK Kernel Reference (Website)
MSLK has organized Attention, GEMM, and MoE kernel usage in one place.
Ramp SWE-Bench (3 min read)
A short article introducing Ramp SWE-Bench has been published.
smevals: An Evaluation Framework for Small and Large Models
smevals provides a framework for evaluating small and large models along with the execution harness.
Microsoft tests new MAI Realtime voice model
Microsoft is testing MAI Realtime, a full-duplex real-time voice model, in MAI Playground.
Teaching Open Models Scientific Research
Open models were trained with RL to improve scientific tool use and biological reasoning performance.
Judge Questions Anthropic's Supply Chain Risk Designation
A judge pointed out that there is insufficient basis to designate Anthropic as a supply chain risk.
NVIDIA Exemplar Cloud: How to Get the Most Out of AI Infrastructure Performance
NVIDIA has released four diagnostic case studies that close the performance gap in AI clusters.
Superlogical website
Superlogical is building a terminal multiplexer that connects interactive, automated, and production work.
Hugging Face AI Storage Buckets
Hugging Face offers AI-dedicated storage with built-in Xet and CDN.
Moonshot's Free Kimi K3 Is Reshaping China's Sovereign AI Strategy
Moonshot released the high-performance open model Kimi K3 for free, shaking up the cost structure of sovereign AI.
Agent Behavior (website)
Defines repeated AI agent behaviors as Markdown specs, used for review and evaluation.
Sessions You Can't Take With You
As Inference APIs bind the core state of sessions to the provider, data ownership and portability are being weakened.
WASTE Inference Engine
WASTE ran the 2.78 trillion-parameter Kimi K3 on 64GB of memory.
Open-Weight LLMs Catch Up in Accuracy
In the ClinReg evaluation, Open-Weight LLMs showed accuracy comparable to closed-source models.
The Agent Graveyard Is No Longer Reality
Enterprise adoption of AI agents has moved beyond experimentation into actual operations and purchasing.