OpenAI launches ChatGPT Voice desktop app

#1 𝕏

OpenAI launched ChatGPT Voice in its desktop app—powered by GPT-Live—to let you control your computer and direct multiple agents in ChatGPT Work or Codex using just your voice.

#2 📝 OpenAI News

Launching Health in ChatGPT - OpenAI announced "Health in ChatGPT" on July 23, 2026, introducing a new product experience that brings health-related capabilities into ChatGPT. The post describes the launch, positioning it as a product offering aimed at helping users with health information and support.

#3 𝕏

Google DeepMind launched the Gemini 3.5 Flash Cyber model and is rolling it out via a limited-access pilot for governments and trusted partners to ensure responsible deployment.

#4 📝 Claude Code Blog

Think through hard problems in voice mode - Product announcement introducing a voice mode that helps users tackle difficult problems using Claude, emphasizing improved productivity and hands-free interaction. The feature is positioned for Claude apps and productivity use cases.

Also covered by: @Claude – Anthropic

#5 𝕏

Qwen launched Qwen-Audio-3.0-TTS with Flash (real-time) and Plus (high-quality) models, featuring inline tags ([whisper], [angry], [breaths], [laughs]), natural-language steering, 16-language support, noise-robust output and one-pass 3-min long-form.

#6 𝕏

Mustafa Suleyman launched MAI-Image-2.5-Pro in Foundry preview, the highest-fidelity, professional-grade image model for ultra-high-quality visuals, detailed editing, and precise in-image text.

#7 𝕏

Mustafa Suleyman launched MAI-Voice-2-Flash, a public preview voice model that’s 2× faster and 32% cheaper at $15 per 1M characters. It’s powering Dynamics 365 Contact Center and cutting GPU costs by up to 89%.

#8 𝕏

Andrew Ng announced OpenWorker, an open-source Mac agent (Windows soon) that automates polished deliverables—customer briefs, Slack messages, calendar updates—across your files and tools. It’s model-agnostic (GPT-5.6 Sol, Claude Fable, Gemini 3.

#9 𝕏

Dharmesh Shah celebrates HubSpot’s public beta launch of Agent Hub and Agent Builder, a toolkit that lets you build custom chat-style AI agents or agentic workflows by mixing your data, tools, and prompts.

#10 𝕏

Philipp Schmid: LangChain now lets you trace GoogleDeepMind Gemini Live speech-to-speech loops in real time—featuring speaker callback hooks, a color-coded live timeline (user in blue, Gemini in orange, tools in custom colors), and separate audio vs.

#11 𝕏

Santiago launched a pay-as-you-go Apify Store skill that lets agents autonomously discover, trigger HTTP 402 payment prompts, authorize USDC on Base, and execute Actors end-to-end—bringing a full AI tools marketplace into agentic workflows.

#12 𝕏

LlamaIndex 🦙 launched new Go and Java SDKs plus a CLI for parsing documents in just a few lines of code or straight from your terminal—grab an API key and start for free today.

#13 𝕏

Cognition built DeepWiki, a free tool indexing over 500,000 repos to help you explore unfamiliar libraries, and will reveal its under-the-hood design in an upcoming LangChain + @jacobtpl meetup.

#14 𝕏

Guillermo Rauch announces that AI Gateway now supports real-time streaming transcription via the new streamTranscribe endpoint in the AI SDK, enabling developers to build live voice agents with unprecedented speed.

#15 𝕏

Boris Cherny uses Fable’s dynamic workflows and a profiler to iteratively tune his code until the p95 latency drops below 300 ms.

#16 in

Colin Matthews suggests kickstarting AI email writing by first defining a clear “good email” rubric—using an LLM to extract criteria from sample emails—and then iterating on drafts against that rubric rather than endless ad-hoc edits.

#17 𝕏

NVIDIA AI ran a hosted RL loop on PrimeIntellect Lab to boost Nemotron 3 Nano’s math accuracy from 22% to 91% for under $5, yielding a downloadable LoRA adapter. The same workflow scales to Nemotron 3 Super and Ultra with just one line change.

#18 📝 Mario Zechner

advanced-context-engineering-for-coding-agents/wsff.md at main · humanlayer/advanced-context-engineering-for-coding-agents - StrongDM has promoted a "lights-off" software factory and OpenAI's Ryan Lopopolo has described their harness-engineering system Symphony, even as companies report outages and codebases degrading. Faros AI found that after AI coding tool adoption in Jan–Feb, PR review quality dropped (+25% more review comments, +22.7% longer comments, 31.3% of PRs skipped review) and production problems rose (incidents per PR +242.7%, monthly incidents +57.9%, bugs per developer +54%).

#19 📝 Mario Zechner

Prompt Caching In Agents | EARENDIL - Transformers retain per-token, per-layer key and value tensors (the KV cache) so subsequent requests can reuse a matching prefix instead of recomputing the whole prompt, and with various tricks those KV caches can be reduced to a handful of gigabytes even for long conversations. Providers either use session‑affinity routing (keeping KV on the same GPU — fast but fragile to overload/restart/eviction) or distribute KV blocks across memory tiers (improves scheduling and recovery but adds movement/indexing complexity), and session trees with forks/rewinds mean prefix overlap and cache hits are often partial and brittle.

#20 𝕏

Sebastian Raschka notes soft distillation transfers richer logit information—ideal within the same model family due to shared tokenizers—but becomes tricky when vocabularies differ. Hard distillation, by contrast, is more straightforward and broadly applicable.

#21 𝕏

Madhu Guru highlights the IAM challenge of managing effectively infinite AI agents spawned by employees—do they inherit their creator’s permissions, what are their lifecycles, and how can we audit them?

#22 𝕏

Madhu Guru explains that Chinese-trained LLMs with open weights can be downloaded and run in your own cloud environment, so your data stays local and the model trainer no longer has access.

#23 𝕏

claire vo is retiring her keyboard by handing her computer to GPT-5.6–powered Codex agents, demoing automated front-end bug detection (03:49), synthetic browser persona research (24:26), LinkedIn inbox cleanup (41:19) and personal shopping (44:39).

#24 📝 Ampcode Chronicle

Event Driven Orbs - Amp orbs can now be woken by external HTTP requests: amp.createWebhook registers a durable webhook endpoint that verifies signatures, deduplicates deliveries, and spawns read-only orb threads with trusted repository/event/actor metadata so the orb can inspect and act on GitHub issues, CI failures, Linear issues, Discord messages, etc. The webhook handler is ordinary TypeScript with the Plugin API (able to append to the owning thread, create new orb threads, maintain durable state, call external APIs, or stop listening), the webhook URL stays the same across plugin reloads and orb restarts and must be kept private, and Amp will connect GitHub webhooks automatically if the orb’s token can administer them otherwise it provides a manual setup step without exposing secrets.

#25 𝕏

Teresa Torres discusses how UK-based Hertility built two AI diagnostic tools for women’s health—most notably a Bayesian network trained on 7 years of linked symptom, hormone test and ultrasound data from over 1 million women—to generate probability-based diagnoses spanning me...