OpenAI shipped two health-focused ChatGPT updates — including a rare-disease diagnosis tool for physicians — while DeepSeek quietly rolled out vision capabilities to its chat interface.
OpenAI published two health-focused updates today: improved health intelligence inside ChatGPT for general users, and a separate tool aimed at helping physicians diagnose rare genetic diseases in children. If you are building anything in health, wellness, or clinical support, these are worth watching closely — they signal where OpenAI is placing its near-term product bets.
DeepSeek introduced vision support in its chat interface at chat.deepseek.com. The Hacker News thread hit 464 points, which is a reliable signal that developers noticed. You can now upload images and ask questions about them directly in the DeepSeek chat UI. No new model name was announced — this appears to be a capability rollout to the existing interface rather than a separate model launch.
OpenAI's health intelligence update to ChatGPT covers better answers to medical questions for everyday users. The rare-disease tool is more specific: it is aimed at physicians trying to identify rare genetic conditions in children, a notoriously hard diagnostic problem where pattern-matching across large datasets can genuinely help. OpenAI published both as separate blog posts on the same day, suggesting a coordinated health push.
Separately, Wired reported on SK Telecom and Anthropic's Mythos export-controls controversy. SK Telecom is a Korean telecom giant that has been investing in Anthropic. The story centers on whether Mythos — an Anthropic initiative — runs into US export control rules when partnering with non-US entities. This is a slow-burn regulatory story, not an immediate builder concern, but worth tracking if you are building on Anthropic APIs for international clients.
ComfyUI hit v0.25.1. If you run image generation workflows locally, this is a routine but real update — ComfyUI is the node-based interface most serious image builders use. Check the GitHub release notes before upgrading if you have custom nodes, as point releases sometimes break compatibility.
Unsloth released v0.1.47-beta. Unsloth is the fine-tuning library that makes running LoRA and QLoRA on consumer GPUs practical — it cuts memory use significantly compared to vanilla Hugging Face training. Beta tag means test before you rely on it for production fine-tuning jobs.
Crawl4AI hit v0.9.0. This is the open-source web scraping library built for feeding LLMs — it outputs clean markdown rather than raw HTML, which saves you a preprocessing step. Version 0.9.0 is a significant minor bump; check the changelog for breaking changes if you have existing pipelines.
Hugging Face published a benchmark post asking whether open models are actually good enough for agentic tasks when you plug in your own tools — not just standard benchmarks. The answer, predictably, is 'it depends on the tool and the model.' Worth reading if you are deciding which open model to use as the backbone of an agent.
Also from Hugging Face: a post comparing fine-tuning techniques beyond LoRA. LoRA is the default everyone reaches for, but the post tests alternatives and asks whether any of them actually beat it on practical tasks. Useful if you are about to start a fine-tuning project and want to know if you should bother looking past LoRA.
The ServiceNow team posted MosaicLeaks on Hugging Face — a study of whether research agents leak confidential information when given access to private documents. Short answer: yes, they can, and the failure modes are subtle. If you are building any agent that touches private data, this is required reading.
1. Test DeepSeek vision for free at chat.deepseek.com — upload a product screenshot or a diagram and ask it to describe what needs fixing. Compare the output quality to GPT-4o vision on the same image. Takes 20 minutes and costs nothing.
2. Set up a Crawl4AI v0.9.0 scraper that pulls a competitor's pricing page weekly and feeds the clean markdown into a ChatGPT prompt that summarizes what changed. No coding background needed if you follow the Crawl4AI docs — the output is plain text your LLM can read directly.
3. Read the Hugging Face MosaicLeaks post before you connect any agent to a folder of private documents. Then add a simple system-prompt rule that tells the agent never to quote raw document text verbatim in its output — a basic but effective first guard.
Confirmed: OpenAI published two separate health blog posts on June 19 — one on general health intelligence in ChatGPT, one on rare childhood disease diagnosis for physicians.
Confirmed: DeepSeek vision is live at chat.deepseek.com and reached 464 HN points, confirming broad developer attention.
Not independently verified by CBW: We have not tested OpenAI's rare-disease physician tool or the updated health intelligence features hands-on. We do not know whether these are available to all ChatGPT tiers or only Plus/Enterprise.
Not independently verified by CBW: The SK Telecom / Anthropic Mythos export-controls story comes from Wired. CBW has not reviewed the underlying documents or legal filings.
Worth noting: ComfyUI v0.25.1, Unsloth v0.1.47-beta, and Crawl4AI v0.9.0 are confirmed GitHub releases, but CBW has not tested any of them. Unsloth is explicitly beta — do not use it in production fine-tuning without testing.
Source: OpenAI blog — rare childhood disease diagnosis — https://openai.com/index/diagnose-rare-childhood-diseases
Source: OpenAI blog — health intelligence in ChatGPT — https://openai.com/index/improving-health-intelligence-in-chatgpt
Source: DeepSeek vision — Hacker News — https://chat.deepseek.com/
Source: Wired — SK Telecom and Anthropic Mythos export controls — https://www.wired.com/story/sk-telecom-anthropic-mythos-export-controls/
Source: Hugging Face blog — MosaicLeaks research agent secrecy — https://huggingface.co/blog/ServiceNow/mosaicleaks
Source: Hugging Face blog — agentic benchmarking on open models — https://huggingface.co/blog/is-it-agentic-enough
Get tomorrow's best AI project in your email. With a guide that works. Free. No spam.