Hi, I'm Onkar. PhD chemical engineer, now building AI apps for people who love food and drink. 🍸 By the Glass Stories: personalized cocktail picks, food and drink pairings, and bar recommendations 🌍 Palate Atlas: food travel recommendations Here I'll share honest takes on the AI models I use every day, plus cocktail, pairing, and food travel tips. The goal is that you learn something new each time. bytheglassstories.app

Today in AI, based on my feed 🧵 • @AIatMeta published a tech report on Muse Glimmer, its 30B open-weight agent model that runs on one consumer GPU • Meta also showed Prometheus, the gigawatt-scale Ohio cluster that trained Muse Wonder • @OpenAI says its latest models are much better at building and migrating COBOL, Fortran and VB code • Leaks: OpenAI's "aurora" alphas look like image models, and "GPT-6.1 Terra" reportedly showed up in eval configs • Claude Fable 5.5: a second checkpoint is said to be with partners. Still not out. Which drops first, Fable 5.5 or Astra? 👀

Next week in AI models, based on my feed 👀 🔒 Closed • Claude Fable 5.5: spotted as "Generally Available" in Azure's catalog (leak) • OpenAI: new aurora-alpha model names spotted, plus GPT-6.1 Astra talk (leak/rumor) • Grok 4.8: lots of chatter, nothing official (rumor) 🔓 Open source • DeepSeek V4.1 Pro: reportedly in testing (rumor) • Kimi K3.1: spotted in testing (leak) • GLM-5.5 and Qwen 4: "coming very soon" (rumor) Which one actually ships first? 🤔 @claudeai @OpenAI @deepseek_ai @Kimi_Moonshot

The Wine Series: American Classics 🍷 California Reds California Whites Oregon & Washington Swipe for the story, the map, the grapes and what to pour with dinner. Save it for your next bottle. Made with Opus 5.5. Directed by Onkar / By the Glass Stories. instagram.com/by_the_glass_s… 🎥 Made with AI. 21+ only. Please drink responsibly. #AmericanWine #WineSeries #ByTheGlassStories #CaliforniaWine #OregonWine #WashingtonWine #NapaValley #WillametteValley #WinePairing #WineLovers

All Claude shipped this week feels like magic. Opus 5.5 in Claude Code responds faster, handles parallel sessions smoothly, and gets more done per prompt. Lately it just keeps getting more effective. Anyone else feeling it? @claudeai @ClaudeDevs @bcherny

Other AI news this week, based on my feed: • Anthropic and OpenAI each put $5M into a Simons Foundation push to help mathematicians work with AI • Gray Swan is running a $300K prompt-injection challenge, backed by Anthropic, Meta, Google DeepMind and Reflection • Anthropic went from about 2,500 merged code changes a day in May to 30,000+ in September, and says that outgrew dev tools built for humans • A new wave of small "decision models" is out: Liquid's d1-ultra and ScaleDown's models give fast yes/no answers and labels instead of chat • Closed models still lead open ones by a wide margin on the Artificial Analysis index What stood out to you?

OpenAI's week, based on my feed: • GPT-Live-1.1 brings new voices, better accents and less "checking..." to Voice in ChatGPT, Codex and dots • GPT-6 Pro and Astra Pro are faster, with multiple agents working in parallel • ChatGPT can now use your Chrome tabs and do tasks for you • Your dot can now be an iOS home screen widget • Fewer "model at capacity" errors, plus another usage reset! Still waiting for voice right inside the chat and a standalone dots app. @OpenAI @ChatGPT @thsottiaux

Claude's week, based on my feed: • Claude Code desktop is built for running lots of sessions side by side • Code Review is now GA for Team & Enterprise, with Quick, Standard, and Deep levels • Faster responses: Claude Code compresses each turn, and cloud sessions start as you type • New defenses against distillation attacks • Mythos 5 designed quantum circuits that beat the best published ones Still waiting for Code Review and a usage dashboard to come to individual plans like Max. @claudeai @ClaudeDevs @bcherny

AI models this week, based on my feed: Closed • Gemini 4 Argon goes global, now in Pro on Antigravity • Meta's Muse Wonder + Wonder Full, built for agents and coding Open • Meta Muse Spark Open (816B, runs on Mac Studios) • Reflection Beam 1.0, Apache 2.0 • StepFun Step 5 Preview weights • Ant Group Ling-3.1-flash • Liquid LFM2.5-24B for local agents Which one are you trying first?

Question for the builders here: I share what I'm learning about AI models, coding agents, and shipping apps. What's most useful to you? Model comparisons, real workflows, build breakdowns, or lessons from what failed? And what's your best advice for getting valuable content to the people who need it?

Switching Opus 5.5 in Claude Code to medium/high effort changed everything for me. I'm getting way more done: data analysis, writing, making videos, and breaking down new topics I'm learning. It's not just a coding tool anymore. @claudeai @bcherny

The test, 3 trials each: every list_pages call site as path:line plus its function. Direct: 10,765 to 24,059 characters into the main context. Delegated: one Agent call, 1,155 back; the subagent's window held ~60,000 tokens. Do this: delegate search to Explore; ask for file:line.