Google DeepMind launches Gemini 3.5 Flash for low-latency inference

#1 📝 OpenAI News

An OpenAI model has disproved a central conjecture in discrete geometry - An internal OpenAI model produced a proof that disproves Erdős’s conjectured upper bound u(n)=n^{1+o(1)} for the planar unit-distance problem by constructing, for infinitely many n, point sets with at least n^{1+δ} unit-distance pairs (the original proof didn’t give δ but a refinement by Will Sawin shows δ = 0.014); the proof — using ideas from algebraic number theory — has been checked and endorsed by external mathematicians.

Also covered by: @Sam Altman, @OpenAI

#2 𝕏

Google DeepMind launched Gemini 3.5 Flash, an optimized edition of its large language model engineered for faster, low-latency inference.

Also covered by: @Demis Hassabis

#3 𝕏

Google AI launched Gemini for Science, a suite of AI-powered tools and experiments to accelerate scientific discovery by connecting and analyzing massive datasets. It’s designed to scale research workflows, speeding up hypothesis testing and insight generation.

#4 𝕏

LlamaIndex 🦙 built a 600-line Next.js demo agent using LiteParse (no vector DB) to ingest SEC filings and answer questions with exact citations highlighted on the original PDF pages. It tackles the ~70% of analysts’ time currently spent pulling numbers from PDFs.

#5 𝕏

Summary: OpenAI highlights AI’s emerging skill in sustaining complex, cross-domain reasoning to accelerate breakthroughs in biology, physics, engineering and medicine.

#6 𝕏

Cursor launched multi-repository support for automations, letting agents reason across codebases to automatically execute, test, and verify tasks.

#7 𝕏

Philipp Schmid launched an API that spins up an isolated Linux sandbox for Gemini to reason, run code, browse the web, and manage files in a single call.

#8 𝕏

Harrison Chase launched Code Interpreter, a lightweight code execution environment for running RLMs and programmatic tool calls (and more) without spinning up a full sandbox.

#9 𝕏

Santiago presents AG-UI, an open, event-based protocol that standardizes agent↔user communication for long-running, nondeterministic tasks—eliminating the need for custom glue code.

#10 𝕏

Teresa Torres outlines how to build Claude-powered AI agents—defining identity, scheduler, tasks, and scripts—to automate prep work, follow-ups, and weekly reviews on custom schedules.

#11 📝 PromptLayer Blog

Best prompt management platforms — Features, comparisons, and recommendations - As teams move from experimental prompting to production-grade AI, they face an infrastructure gap managing prompt versions, models, environments, and changes. The article outlines that gap and compares platform features to help teams choose.

#12 𝕏

Andrew Ng launched a short course with Google Cloud on building self-evaluating AI agents for image and video generation, teaching three evaluation techniques—image-text similarity scoring, LLM judges for custom criteria, and structured rubrics.

#13 𝕏

Julien Chaumond launched Hugging Face Hardware, a community-driven dashboard revealing the real-world GPUs & CPUs powering open-source AI, plus VRAM distribution and inference hardware trends.

#14 𝕏

Sebastian Raschka flags a new LLM parallel block design that matches vanilla transformer performance while delivering significantly higher throughput.

#15 in

Guillermo Rauch launched the AI Gateway plugin for WordPress, bringing any AI model or provider—covering text, image, video, and audio—to 42% of the web.

#16 𝕏

claire vo 🖤 notes that Anthropic has locked down enterprises on Claude en masse, but warns vendor lock-in slows you from seeing the real frontier. Cutting-edge builders instead hop between OpenAI’s Codex, CoWork, and AI Studio to stay fast, flexible, and impactful.

#17 𝕏

clem 🤗 celebrates Cohere’s release of the Apache 2.0-licensed “command-a-plus-05-2026-bf16” model on Hugging Face, highlighting their strong open-source momentum.

#18 𝕏

clem 🤗 built and released Carbon: a frontier DNA base model with open weights, training code, and data pipeline. It’s 275× faster than the next-best model, runs locally on a laptop, and can process a whole human genome on a single GPU.

#19 𝕏

claire vo 🖤 recaps her favorite #GoogleIO launches—Antigravity, Gemini, AI Studio, Flow, Omni, Stitch, Pomelli, and more—for engineers and designers. She also shares hands-on trials (and occasional failures) with the new tools.

#20 𝕏

Peter Yang stresses that teams should “just try a lot and build to learn,” running 3–4 rapid iterations to discover what works, and stick to 90–120-day roadmaps instead of year-long plans.

#21 𝕏

Sam Altman spotlights AGI’s three key impacts—accelerating research, powering companies, and personal AI—celebrates the “unit distance” breakthrough and offers $2M in OpenAI credits to every YC company. He urges ramping up personal AGI to help individuals achieve their goals.

Also covered by: @Sam Altman, @OpenAI

#22 𝕏

Logan Kilpatrick calls Gemini 3.5 the start of a new era after 2½ years of building its infrastructure, products, and team, and emphasizes that “the model is the product,” urging users to keep the feedback coming.

#23 𝕏

Demis Hassabis unveils Gemini 3.5 Flash, a compact LLM using Flash Attention for sub-second inference and reduced GPU memory footprint, now available on Google Cloud’s Vertex AI.

Also covered by: @Demis Hassabis

#24 📝 Surge AI Blog

Slop is a choice. Introducing Antidote. - Antidote is an evaluation framework that emphasizes expert human reviewers who read and grade AI outputs to push model evaluation beyond superficial or automated metrics. Its goal is to reduce low-quality "slop" by relying on human judgment and nuance.

#25 𝕏

Boris Cherny revamped the usage UI so you can now run `/usage` to see exactly which calls are consuming your tokens.