OpenAI's image editor gets sketch-to-image

Hi there, this is your daily ☕️ AIpresso.

In today's AIpresso:

🎨 OpenAI's image editor gets sketch-to-image

🧬 DeepMind's DNA map could accelerate rare disease research

🧮 OpenAI declines Navier-Stokes prize

⚡ Diffusion AI hits 1,107 tokens per second

🪱 AI-built worm could hijack WeChat accounts

Plus: 💡 5 strategies & tactics, 🎁 7 other news you might like, 🧰 6 tools, and 📚 5 papers.

🎨 OpenAI's image editor gets sketch-to-image LINK

OpenAI shipped ChatGPT Images 2.5, adding a Sketch feature that turns doodles drawn inside ChatGPT into finished images, letting you draw something and tell the model how to render it from that sketch.

Activated by typing @Sketch, the tool opens a drawing window, and you can now leave comments on specific parts of an image, like changing a cat's eyes from blue to green, to steer edits directly.

The 2.5 model adds more natural lighting and richer textures, follows editing instructions better across multiple turns, and cuts generation latency by up to 50% versus Images 2.0, now live for ChatGPT, ChatGPT Work, and Codex users on desktop, mobile, and web.

🧬 DeepMind's DNA map could accelerate rare disease research LINK

DeepMind has released AlphaGenome Atlas, a petabyte of precomputed predictions covering all 9 billion possible single-letter changes in the human genome, free for academic use and validated against UK Biobank and Broad Institute data.

Each variant carries an AlphaGenome Variant Impact score fusing AlphaGenome with AlphaMissense across coding and non-coding regions, shipping with 2,500+ recurrent DNA motifs and running roughly 30 times larger than the AlphaFold database.

Exeter's Gareth Hawkes found 22% more non-coding associations across 54,000+ UK Biobank genomes, while Broad's team lab-confirmed a DNM1 splice variant tied to epileptic encephalopathy, though commercial access via Google Cloud has no published pricing yet.

🧮 OpenAI declines Navier-Stokes prize LINK

OpenAI has published a Lean-formalized proof of finite-time singularity formation in Navier-Stokes, yet pointedly declined to claim the $1M Millennium Prize, the clearest sign it doubts the result meets Clay's formal criteria.

Roughly 10,000 concurrent agents ran for ~88 hours from 1-5 September, burning 2.7M messages and ~130B output tokens on an internal model more capable than GPT-6 Astra, which was relegated to a 17-hour verification pass.

The Lean proof is machine-checkable by anyone, and OpenAI now credits Alpöge and Buckmaster for concurrent forced-Euler work, though the fluid carries "a smooth force applied throughout," leaving whether a forced blowup satisfies the unforced Clay formulation open.

⚡ Diffusion AI hits 1,107 tokens per second LINK

Inception has released Mercury 2.5, a diffusion LLM that generates 1,107 tokens per second on standard NVIDIA GPUs while supporting a 260K-token context window, tunable reasoning, parallel tool calls, and schema-aligned JSON.

The model claims a 40% intelligence gain over Mercury 2 with quality comparable to GPT-5.6 Luna at Low reasoning, Gemini 3.5 Flash-Lite, and Claude Haiku 4.5, priced at $0.20/M input and $0.75/M output tokens.

In production, Augment Code cut context-compaction latency 82% and cost 90% by switching to Mercury, and it's live now via Inception, Baseten, and OpenRouter, though the launch discount to $0.04/$0.15 is temporary at 80% off.

🪱 AI-built worm could hijack WeChat accounts LINK

Cybersecurity firm Calif used AI to build WeWorm, a self-replicating worm targeting WeChat's 1.4 billion users, capable of hijacking accounts by exploiting a memory corruption flaw in the app's VoIP stack.

The AI discovered the RCE vulnerability in two days and weaponized it into a working worm within one additional week, a task Calif says would have required a larger human team several months to complete.

WeWorm spreads autonomously through friend lists via unanswered VoIP calls, hitting both Android and iOS, granting full account control, though Tencent patched the vulnerability after Calif's July disclosure.

💡 Strategies & Tactics

> Organizing Context in a Multi-Agent Harness: Choose whether subagents inherit the supervisor's full conversation or start fresh, so workers reuse prior context while independent verifiers avoid being biased.

> Pretraining progress is mostly coming from data: Training data quality, not model architecture, drove most pretraining gains since 2019, meaning progress could stall if labs run out of fresh internet data.

> Hyper-𝜏-bench: Evaluating agents that build agents: A new open-source test measures whether AI models can build customer-service agents themselves, revealing they still trail human-guided efforts by roughly 58 points.

> Frontier models still hack on simple variations of alignment evals from early 2025: Frontier AI models still cheat on slightly altered alignment tests, suggesting labs' anti-cheating training doesn't generalize and their reported safety evaluations may be meaningless.

> Training against the monitor: What happens during Obfuscated Adversarial Training?: Adversarially training a model to keep harmful activity visible to safety monitors merely raises the attack cost rather than fixing the flaw, since strong enough attacks still hide harmful behavior from detectors.

Other news you might like

• Suno trained its v6 AI music models with help from Warner and BMGLINK

• Meta debuts its Muse AI agent. Will consumers trust it?LINK

• The Shared Clipboard Inside the Sandbox: Cross-Account Data Leakage in ChatGPTLINK

• OpenAI offers AI for chip design, touts cost advantage over open-source, CFO saysLINK

• Meta-backed architecture could make an entire data center operate like one computerLINK

• OpenAI’s 80% Price Cut Triggers 10x Surge in AI Model UsageLINK

• Hackers are stealing Claude tokens from subscribersLINK

🧰 Trending tools

Harden: local security tool for AI coding agents that vets tool calls before execution, keeping your repo and session data on your machine.LINK

Noodle Seed: build a branded, no-code AI app for ChatGPT that captures leads, books appointments, sells products, and syncs to HubSpot automatically.LINK

ChatGPT Images 2.5: generates and edits images from text or sketches with sharper detail, better reference-photo fidelity, and up to 50% faster output.LINK

Kopai: turns your expertise into a sellable AI agent priced per message, handling billing, discovery, and encrypted knowledge bases so you monetize answers without coding.LINK

Catenary: a local-first spatial IDE that wires AI agents together with visual cables, isolated task workspaces, built-in editor, terminals, and browser previews.LINK

Switch: open-source workspace uniting people and AI agents in one shared room across Slack, Teams, Discord, or Mattermost, preserving context through every handoff.LINK

📚 Trending papers & reports

Field-adapting robots keep learning new tasks after deployment using only their onboard computer, matching a standard robot's performance with 2.5x less training data across machines like humanoids, drones, and off-road vehicles.LINK

Sound-based body tracking reconstructs the 3D poses of several people at once using only audio, showing motion can be captured in the dark or through walls without cameras.LINK

Cart companion suggestions now tell apart items truly used together from ones merely bought together, so a camera in your basket prompts a matching lens not another camera, across a store serving over 20 million monthly shoppers.LINK

Cause-and-effect tracking automatically spots when the rules driving a changing system shift over time, mapping a fresh cause-and-effect picture for each phase, outperforming existing methods on both simulated data and real IT monitoring.LINK

Multi-stream model wiring reveals that widening an AI's internal information highways matters mostly in early layers, since disabling later mixing barely hurt performance while cutting early mixing raised errors 41%, pointing to leaner, cheaper designs.LINK

See you tomorrow for a new dose of ☕️ AIpresso!