AI Briefing
TLDR AI news
The latest AI and developer news from TLDR AI, with short summaries.
Visit website →Perplexity Proposes JOLT: Joint On-Policy Learning and Teaching for Efficient RL Post-Training
JOLT jointly trains a single model as both teacher and student to provide dense, aligned guidance, achieving higher accuracy and significantly better sample efficiency than GRPO on math, coding, and agentic benchmarks.
WorldTUI: 360M-Parameter Model Simulates Terminal Interface Without Shell Backend
Fine-tuned on 2B tokens of synthetic data, the model predicts screen deltas from keypresses, outperforming prompted baselines on shell and file readback tasks.
PickOpposable runs 2B GUI-Owl 1.5 model on Android for on-device app control
The app uses GUI-Owl 1.5 2B to operate Android apps locally, achieving 7.3s step times on M2 CPU and 236ms vision encoding on Galaxy S25 NPU.
Ant Group and vLLM Community Open-Source Humming for High-Performance Quantized Inference
The library supports NVIDIA Turing and later architectures, achieving significant speedups over vLLM Marlin and DeepGEMM across dense and MoE workloads.
SparkLLM Team Releases Spark-X2.6-4B, a 4B On-Device Model Matching 9B Performance
The model supports a 1,048,576-token context and offers controllable thinking budgets to balance reasoning quality and latency.
PickElevenLabs launches Eleven Realtime v1 Mini and Max speech-to-speech models
The new models combine speech recognition, turn-taking, LLM, and voice into a single end-to-end system available via ElevenAgents and an OpenAI-compatible API.
Ramp Builds Laminar, an In-House Managed Agents Platform
Ramp developed Laminar to unify its agentic products on a model-agnostic, cloud-native stack, replacing fragmented vendor solutions and open-source frameworks.
PickDeepSeek Open-Sources DeepInfer Framework for V4.1 Flash Inference
The open-source DeepInfer framework supports NVIDIA and Huawei Ascend platforms with a single-cluster throughput of 155 trillion tokens per day.
llm-d 1.0 releases production-ready LLM inference stack with multi-tenant QoS and stable API
The open-source Kubernetes-native stack supports vLLM and SGLang across heterogeneous hardware, claiming up to 3x higher output throughput and 2x faster TTFT.
OpenAI adds PIN and biometric locking for ChatGPT conversations
The feature moves chats to a dedicated section with separate memory, available on most plans but not Healthcare.
Databricks MemCurrent Improves Agent Pass Rate 3x and Cuts Token Usage by 31%
The architecture combines episodic and derived memory to achieve 68.4% pass rates on internal benchmarks while reducing latency by 26%.
a16z Analysis: Superhuman Hardware Design Is Easy, But Physical Manufacturing Remains the Bottleneck
The author argues that while frontier models will soon achieve superhuman CAD and PCB design, the lossy interface between digital bits and physical atoms creates distinct barriers in manufacturing, testing, and data collection.
Pichai: Google Sacrifices Cloud Revenue to Fund Frontier AI, Argon Outperforms Rivals
Gemini 4 Argon processed 25 trillion tokens in 72 hours, tripling Gemini 3.0 usage, while Pichai confirms compute constraints are driving efficiency breakthroughs.
Meta's Muse hits 10.9M downloads in 40 days, outpacing ChatGPT in US retention
Sensor Tower data shows Muse retains 30% of users after 7 days, significantly higher than the 19% average for competitors like ChatGPT and Claude.
Paper Proposes Pair-Matched HodgeRank and Post-Control Risk to Complement g-mean² Monitorability Metrics
The paper addresses g-mean²'s sensitivity to task mixture and deployment risk by introducing pair-matched comparisons and post-control rejection metrics.
Office Intelligence Research Program Introduces Environments for Long-Running AI Knowledge Work
Office Intelligence is a research program providing environments, evaluation criteria, and training resources for AI agents performing sustained professional knowledge work.
Nscale and Anthropic reportedly in talks for second $45B compute deal
Bloomberg reports the companies are on the verge of announcing a new agreement, following a similar deal struck just over a month ago.
AI Benchmarks: Climbing the Wrong Hills
Contrary Research argues that 'benchmaxxing' is misaligning AI development with real-world utility, citing cases where top-ranked models fail in practical tasks.
Gemini Notebook adds co-creation controls for Slide Decks and Video Overviews
The feature allows users to review outlines and producer notes before rendering, rolling out globally to all users over the upcoming week.
Jack Clark tells IMF AI Singularity has begun, citing 98% score on unpublished physics problems
Anthropic co-founder argues AI now autonomously accelerates scientific discovery, projecting a $44 trillion US GDP by 2030 if adoption scales.