AI Briefing
TLDR AI news
The latest AI and developer news from TLDR AI, with short summaries.
Visit website →When Can LLMs Learn Reasoning Under Weak Supervision
Only Qwen, which has a long pre-saturation phase, succeeds at RLVR generalization under weak supervision, while Llama fails.
Coding agents ignore their own budgets
Even when shown their budget, coding agents couldn't cut their own spending.
Image Generation Prompting Guide (38-min read)
This summarizes practical prompting principles that improve quality, speed, and editing for gpt-image-2.
Qwen3.5-Omni Technical Report
Qwen3.5-Omni achieves SOTA with 256k context and over 100 million hours of audio-visual data.
OpenAI Developing Always-On Agent Platform for ChatGPT
OpenAI is preparing a 24/7 agent platform inside ChatGPT.
ChatGPT Images 2.0
OpenAI unveiled ChatGPT's new image generation feature, 'Images 2.0.'
Claude Can Now Create Live Artifacts in Real Time
Claude generates dashboards and trackers in real time, connected to apps and files.
9 Notable Remarks from Demis Hassabis and Sebastian Mallaby on a San Francisco Stage
From OpenAI's bankruptcy odds to AlphaFold and AGI, 9 key remarks from Demis Hassabis have been revealed.
Microsoft Shifts GitHub Copilot Users to Token-Based Billing and Tightens Rate Limits
GitHub Copilot is moving from request-based to token-based billing.
OpenAI Stargate: Status of US Sites
Stargate is scaling into a massive AI infrastructure network, deploying over 9GW across 7 sites in the US.
Flash Vision-Language-Action Inference for Autonomous Driving: FlashDrive
FlashDrive optimizes the entire autonomous driving VLA inference pipeline, cutting latency from 716ms to 159ms.
DeepMind's TIPSv2 Vision-Language Encoder
TIPSv2 significantly boosts patch-text alignment through iBOT++, Head-only EMA, and multi-granularity captions.
Google Adds Subagents for Parallel Coding to Gemini CLI
Gemini CLI now splits tasks among subagents to process them in parallel.
Even 'uncensored' models can't say what they want to say
Even 'uncensored' models have a flinch that lowers the probability of banned words on their own.
Modular Post-Training
BAR is an approach where domain-specific experts are each post-trained independently and then merged into an MoE.
Boosting Training Efficiency with Effective Training Time
Meta pushed ETT% above 90%, significantly reducing training inefficiency.
Jeff Bezos Nears $10 Billion Investment for AI Startup
Bezos's AI startup Project Prometheus is on the verge of securing $10 billion in funding.
Qwen3.6-Max-Preview: Smarter, More Refined, and Still Evolving
Qwen3.6-Max-Preview significantly boosts coding, knowledge, and instruction-following performance.
Moonshot AI launches Kimi K2.6 in Kimi Chat and API
Moonshot AI has released Kimi K2.6 via Chat and API, touting it as open-source state-of-the-art.
Chronicle – Codex
Chronicle augments Codex memory with on-screen context to enrich task context.