Daily TEA – OpenAI Just Made Coding a Conversation

AMD says agentic work is a CPU problem, ramble coding, GPT-Live lands in Codex, Kimi K3 trades tokens for taste, India goes after Bitchat’s source code

Jul 27, 2026

Hello, dear TEA-mates! Here is what you need to know today.

1. 🖥️ AMD Says the CPU Is Back in the AI Conversation

Speaking to theCUBE at AMD Advancing AI 2026 on July 23, Derek Dicker, corporate vice president of AMD’s Enterprise Business Group, argued that the shift to agentic workloads has changed what enterprises actually need to buy. “If you looked at GPU being the center of a lot of these workloads, what’s happened over time as agentic has unfolded is a realization that it’s, as much a CPU workload as it is a GPU workload,” Dicker said. He framed enterprise demand as driven by three things: silicon choice, data center modernization, and the total cost of tokens. Rather than build competing systems, AMD published an open rack-scale specification built on Meta’s OCP design and rallied partners around it. Its Helios reference design pairs 72 AMD Instinct GPUs with AMD EPYC CPUs and AMD Pensando networking in one unified rack. Dicker also tied openness to sovereignty, noting that healthcare, telecom, and national buyers increasingly want to host AI on their own infrastructure and their own soil. (Read More)

🫖 TEA For Thought: “In the era of AI, especially when it comes to agentic workflows, the CPU is almost as important as the GPU.”

2. 🎙️ How to Become a 10x Ramble Coder

Moe Khalil, CTO at Webhound, makes the case that talking to your coding agent beats typing at it, picking up on Andrej Karpathy’s recent post describing “a nice long ramble session” as his new favorite way to work with LLMs. Khalil’s argument is compression: a typed prompt flattens a messy train of thought into a tidy sentence, and everything useful gets thrown away. He walks through a login-screen example where the raw ramble reveals five things a polished prompt never would, including that his users are non-technical, that he hates redirecting people to a second page, and that he sometimes chases vanity. His tips are to lean into your uncertainties (the option you almost picked reveals your constraints), to lean into your emotions (they expose your priorities), and to try “multiplayer rambling,” where he and his cofounder record hours of deliberate disagreement on a voice memo app. They then transcribe it with OpenAI’s Whisper using speaker identification and feed the transcript into Codex as ground truth. For solo use he points to dictation apps like Wispr Flow, Willow, and Aqua Voice. (Read More)

🫖 TEA For Thought: “Time to speak your mind and build something cool.”

3. 🗣️ OpenAI Puts Full-Duplex Voice Inside Codex

Two weeks after launching GPT-Live on July 8, OpenAI moved the full-duplex audio model into the ChatGPT desktop app on macOS and Windows, wiring it directly into Codex and ChatGPT Work, which together have more than 10 million weekly active users. Full-duplex means the model listens and speaks at the same time, dropping in verbal acknowledgments like “got it” without cutting the user off, while heavy reasoning is handed to background models such as GPT-5.5. Engineers can fire several concurrent task threads from one spoken prompt, for example investigating an authentication bug, reviewing a pending API migration pull request, and generating missing unit tests at once, traced across Slack, GitHub, and local codebases. On macOS, “Appshots” and screen context let the assistant read the frontmost window alongside local files and codebase structure. Build 26.715 adds multi-folder projects and remote execution from the ChatGPT iOS app. Access is limited to paid Plus, Pro, Business, Enterprise, and Education plans, and voice-triggered tasks draw from the same Codex and ChatGPT Work quotas as typed ones. (Read More)

🫖 TEA For Thought: “Jarvis has arrived! Just curious how many tokens this voice mode costs overall.”

4. 🧠 Kimi K3 Trades Tokens for Intelligence

DesignArena reports that Moonshot AI’s Kimi K3 now ranks first on its single-shot Frontend Arena with an Elo of 1392, 10 positions above Kimi K2.6 and 16 above Kimi K2.7 Code, the biggest jump the lab has recorded in the Moonshot line. The cost is reasoning volume: K3 burns over 12 times more thinking tokens than Claude Opus 4.8 and more than double Kimi K2.6. Digging into the traces, DesignArena found K3 runs a full agentic loop inside its own chain of thought, moving from planning to decision-making to designing individual components, and writing sample code mid-thought so the final generation gets the interactions right. It produces over 10 times more code during reasoning than any other model in its line, spending more reasoning tokens writing code than reasoning in prose. K3 also one-shots valid Unsplash image IDs from its learned index of the web, then checks them mentally before committing. The result is slower generations with higher preference scores, setting a new Pareto frontier. Open weights are due by July 27, 2026. (Read More)

🫖 TEA For Thought: “By choosing this strategy for its reasoning traces, Kimi K3 deliberately produces results that take more tokens but perform better, essentially trading tokens for intelligence.”

5. 🛰️ India Targets Bitchat’s Code, Not Its Content

Jack Dorsey posted on X what he said was a notice from India’s Ministry of Home Affairs directing GitHub to restrict access to three repositories for Bitchat, his offline Bluetooth mesh messaging app, within three hours. The notice, dated July 23 and apparently issued by the Indian Cybercrime Coordination Centre, does not point to any specific unlawful post or file. It argues instead that the app’s anonymous, decentralized architecture lets people communicate during internet shutdowns and hampers “lawful interception, attribution, and traceability.” The timing follows weeks of student-led protests in New Delhi over alleged exam paper leaks, during which authorities suspended internet services. Sensor Tower data shared with TechCrunch shows India accounted for roughly 85% of Bitchat’s global downloads between July 17 and July 23, up from about 1% over the prior 30 days, with more than 91,000 downloads in five days and over 330,000 daily active users on Thursday. Digital rights groups including SFLC.in, the Internet Freedom Foundation, and Access Now questioned the legal basis, and the repositories remained accessible in India on Friday. (Read More)

🫖 TEA For Thought: “Freedom of speech is a God-given right. No government or authority can take it away, even in the name of good.”

🛠️ Skill of the Day

The Ramble Decoder: turn a messy, contradictory voice note into a clean brief your AI can actually build from.

You are my thinking partner. Below is a raw, unedited transcript of me
thinking out loud. It rambles, contradicts itself, and changes direction.
Do NOT clean it up into a summary. Instead, decode it.

Give me five sections:

1. WHAT I ACTUALLY WANT: the goal in one sentence, in my words.
2. MY REAL CONSTRAINTS: the limits I revealed accidentally, not the ones
   I stated. Pay special attention to options I considered and rejected,
   and tell me what each rejection says about me.
3. MY PRIORITIES, RANKED: infer these from where I got emphatic,
   frustrated, or repeated myself. Quote the line that proves each one.
4. CONTRADICTIONS: every place I said two incompatible things. For each,
   ask me one short question that would settle it.
5. THE BRIEF: a tight, buildable spec that respects all of the above.

Be honest if I never actually said what I want. Do not invent preferences
I did not express.

TRANSCRIPT:
[PASTE YOUR RAMBLE HERE]

Record a voice note on your phone, dictate it into any AI tool, then paste it in. The messier the ramble, the better this works.