🤖 AI HOT Daily · Jun 25, 2026

OpenAI Bidi 1 bidirectional voice model in testing, Qwen-AgentWorld open-sourced beats GPT-5.4, Gemini 3.5 Flash gets computer use, Doubao Pro launches, OpenAI & Broadcom unveil Jalapeño inference chip, NSA loses Mythos access, DFlash boosts throughput 15x.

1. 1. OpenAI's Biggest Voice Upgrade: Bidi 1 Bidirectional Model

   Bidi 1 bidirectional voice model supports simultaneous speaking and listening, allowing users to interrupt mid-conversation with new instructions. Not yet officially announced; wider testing this week.

2. 2. Qwen-AgentWorld Open-Sourced: Predict Before Acting

   First native language world model covering MCP, Search, Terminal, SWE, Web, OS, Android. Trained on 10M+ real interaction trajectories, surpassing GPT-5.4 and Claude Opus 4.8 on AgentWorldBench.

3. 3. Gemini 3.5 Flash Gains Computer Use Capability

   Google integrates computer use as a built-in tool for Gemini 3.5 Flash, supporting agent operations across browser, mobile, and desktop environments with new adversarial training.

4. 4. Doubao Pro Officially Launched: 500 Yuan/Month

   Based on Doubao 2.1 models, supports local computer control, browser operations, Skills, scheduled tasks, and built-in Office suite. Standard 68 yuan/month, Pro 500 yuan/month.

5. 5. Figma Config 2026: Canvas AI Powered by Third Parties

   Figma expands canvas to code, animation, and 3D. AI features rely on external models from Anthropic, OpenAI, and Google, with inference costs squeezing margins.

6. 6. OpenAI & Broadcom Unveil Jalapeño Custom Inference Chip

   OpenAI and Broadcom jointly unveil Jalapeño, a custom AI chip optimized for large language model inference, targeting improved performance, efficiency, and scale.

7. 7. NSA Loses Mythos Access Amid Dispute with Anthropic

   The NSA has lost access to the Mythos system due to a dispute with Anthropic. Mythos was previously revealed to have breached NSA classified systems within hours.

8. 8. DFlash: Block Diffusion Drafting Achieves 15x Throughput

   UC San Diego proposes DFlash generating whole token blocks in one forward pass. Up to 15x throughput improvement on NVIDIA Blackwell for gpt-oss-120b, ~1.5x EAGLE-3.

9. 9. Engineering: The Most Resilient Job of 2025 Despite AI Fears

   SignalFire data shows engineering is the most resilient job of 2025. Big tech engineering roles only declined 11%, while early-stage startup engineering hiring grew 7% vs 2019.

10. 10. New GPT-5.5 Instant: More Engaging Conversations

    Updated GPT-5.5 Instant better understands intent behind questions and adapts responses accordingly, handling complex constraints more reliably. Rolling out to paid and free users.