TL;DR: No single LLM is optimal for every query and budget. This work unifies the scattered field of LLM routing into "five building blocks" and ships an open-source infrastructure bundling 16+ routers with a dedicated benchmark. Title: LLMRouter: Unified Infrastructure for Developing, Evaluating, and Deploying LLM Routers URL: Key points 🧩 Decomposes any router into 5 components: Context Encoder / Model Encoder / Scoring / Decision Rule / Learning Signal 📊 xRouteBench spans 5 tracks (Generic, Memory, Vision, TimeSeries, Personalized), 4,767 queries total 🛠️ A dense query-model matrix, built by dispatching every query to all 18 candidate models, serves as both training supervision and test bed 🚀 Learned routers beat the strongest fixed-model baseline (always pick the largest) by 14.6% relative ⚖️ No router dominates: RouterDC leads on accuracy but drops to 10th under tight cost constraints 🔁 Multi-turn routing adds cost without consistent gains (Router-R1 at just 22.3%) 👤 User conditioning helps, but the best design flips between simulated personas and real feedback A solid step that proves "the best router depends on task and budget" and lays a foundation for fair comparison. #LLM# #ModelRouting#

TL;DR: Just three weeks after Gemini 3.6 Flash, Google drops 3.7 Flash — major gains across coding and agent benchmarks, at half the price during the intro period. Introducing Gemini 3.7 Flash Highlights: 🚀 DeepSWE v1.1: 65.3% (up from 49.0%, +16.3pt vs 3.6 Flash) 🖥 FrontierCode 1.1: 43.6% (up from 34.4%, +9.2pt) 🌐 WebDev Arena Elo: 1588 (up from 1538, +50) 📄 GDP.pdf document comprehension: 34.0% (up from 22.0%, +12.0pt) 🤖 AutomationBench: 30.4% (up from 17.0%, ~1.8x improvement) 💰 Intro pricing: $0.75 input / $3.75 output per 1M tokens — half the cost of 3.6 Flash 🌍 Available via Gemini API, AI Studio, Android Studio, Gemini Enterprise, and Gemini Spark in 160+ countries 🛡 Updated Frontier Safety safeguards against CBRN and cybersecurity misuse Sharpened for agents and coding, it's already earning "significantly better" feedback from Box, Databricks, Harvey, and LangChain in real-world production workflows. An update at this quality level every three weeks — the Flash release cadence is relentless. #Gemini# #LLM#

TL;DR FoundationStereo, the high-accuracy zero-shot stereo model, sped up by over 10x with almost no accuracy loss. It finally runs in real time. Title: Fast-FoundationStereo: Real-Time Zero-Shot Stereo Matching URL: Points 🎯 Problem: stereo had split into "accurate but heavy foundation models" vs "fast methods that sacrifice accuracy and generalization" 🧩 Divide-and-conquer: each of three bottlenecks accelerated separately (feature extraction, cost filtering, refinement) 📚 Feature extraction: distill hybrid monocular + stereo priors into a single efficient student 🔍 Cost filtering: blockwise NAS over 8 blocks cuts search from O(n^N) to O(n), with ILP minimizing error under a latency budget ✂️ Refinement: structured pruning via ConvGRU recurrent dependency graphs and Taylor-expansion importance 🗂 1.4M in-the-wild stereo pairs auto pseudo-labeled with a normal-consistency check ⚡ Middlebury-Q: 4.51% BP-2 at 49ms (vs 2.64% at 496ms, 10.1x slower), down to 21ms with TensorRT A new SOTA among real-time stereo methods, well suited for robotics and AR deployment. #StereoMatching# #ComputerVision#

TL;DR AI use is shifting from "chat Q&A" to "delegating whole tasks." OpenAI's usage data shows the unit of knowledge work moving from single interactions to long-horizon tasks. 📊 Title: How agents are transforming work URL: Highlights 🧑‍💼 The unit of knowledge work shifts from "interactions" to "delegated, long-horizon tasks" ⏱️ In May 2026, 80.6% of individuals made a Codex request >30 min of human work, 70.2% >1 hour, 25.6% >8 hours 🔀 By June, 99th-percentile users generated 60+ hours of agent turns per day across parallel agents 📈 Non-developer users exploded vs Aug 2025: 137x individuals, 189x organizations 👥 Codex now has 5M+ weekly active users (6x+ since the February desktop launch) 🚀 Fastest-growing tasks: data analysis +110%, research +37%, knowledge artifacts +36% (week over week) 🏢 Even non-technical departments like Legal and Recruiting use Codex as their primary AI tool The move to redesign work around "humans plus many parallel agents" is now visible in the numbers. #AIAgents# #FutureOfWork#

TL;DR Reliable agentic AI comes not from "a better model or prompt" but from explicitly engineering context and the orchestration harness—lessons from PRINCE, a Bayer × Thoughtworks pharma system. 🧪 Title: Building Reliable Agentic AI Systems URL: Highlights 🧭 Sequential agents with pause points: clarify intent → Think & Plan → Researcher → Reflection → Writer, verifying step by step 🔁 Three reflection loops: process (trajectory), data (evidence sufficiency), draft (output completeness) catch distinct failures 🔎 Hybrid retrieval: query expansion n=5, weighting 0.7 semantic + 0.3 keyword, bge-reranker cutting ~20 → 7 chunks 🗃️ Structured data via Text-to-SQL: SELECT-only, up to 3 self-corrections, ≤50 rows per query 🛟 Harness engineering: state persisted in PostgreSQL/DynamoDB, resume from failure point, automatic provider fallback 📌 Sentence-level citations plus RAGAS and Langfuse evals; daily batch checks on live traffic catch hallucinations 🏷️ NER extracts entities from study PDFs; high-confidence fields auto-update, low-confidence quarantined for human review The "even with big context windows, selectivity still matters" stance rings true to anyone shipping this stuff. #AIAgents# #LLMOps#

tl;dr - subagent behaviour working on adding multi chat and subagents to the agents starter (yay!) and I have a curious product direction/question. our subagents can be full fledged chats themselves. which means they could not only be async while they work on their thing and you continue, but you could continue "talking" with them after they've "returned" a result. so what should the default behaviour in the starter be? - readonly, no input. this is what most (all?) products/devtools like this do atm - have chat, but it's only followups, doesn't affect the main chat - add a "send back/summarize to main chat" this feels powerful and underexplored I'll probably ship option 1 for now, but there's something here... anyway, multichat/subagents in starter template coming this week

TL;DR (and this is HUGE): looping makes sense for a select few niche markets again because of a NEW MECHANISM that sustainably compresses borrow rates relative to demand. Despite arguing that Junior tranches are often superior to leveraged loops, there's ONE META that actually makes looping interesting for me again. And if you're not obsessed with DeFi, you probably haven't heard about this. IN SHORT, teams like @fraxfinance and @Paxos give kickbacks equivalent to the TBILL yield for institutional LPs holding or lending frxUSD or USDG. THAT MEANS, big'ol LPs can hold NON-YIELD BEARING assets like PYUSD, USDG, and frxUSD and get 3-4% yield paid to them on a semi-regular basis. On top of that, they can then lend those assets on money markets for the 3-4% base yield PLUS whatever the interest rate pays. But here's the kicker. Because those rates are paid OTC to these lenders, the borrow DOES NOT PAY the additional TBILL yield like you would if you were borrowing a yield-bearing stablecoin. So the borrower might only be paying 4% in interest, despite the collateral depositor making 5-7%, which is very juicy for an institution who wants to be holding USDG or PYUSD or frxUSD, e.g. This means lower interest rates for borrowers, because lenders are no longer depending wholly on interest rates for compensation. Of course, this is unique to specific markets, but I feel fairly confident this meta will drive some insane TVL growth in the near future.