Today's thread
The more work we hand to AI, the bigger two questions get. One is who shares the cost when AI changes people's jobs. The other is which tools let AI work longer, at larger volume, and more accurately. Today's four stories each answer one of these.
Anthropic plans to set aside 5% of its shares for a public benefit fund and explicitly mentions AI-driven labor market change as a direction for support. On the open-source side, herdr, Lightpanda, and PaddleOCR-VL 1.6 each tackle a bottleneck that agents (AI workers that chain several steps on their own instead of a person) hit: sessions dropping, the cost of reading the web, and numbers in PDFs getting garbled.
Anthropic puts 5% of shares into a public benefit fund — pending approval
On October 18, Anthropic announced the Anthropic Public Benefit Fund and committed 5% of its shares to it. The commitment is subject to stockholder approval. The board may add another 2% for every 50% increase in the company's value above a set baseline, but the baseline has not been disclosed.
The fund will be a wholly owned LLC rather than a charity. Anthropic says this keeps options open, from donations to traditional nonprofits to direct distributions to specific populations, and the equity cannot be used for lobbying or political giving. Combined with the co-founders' pledge at founding to donate 80% of their equity and an employee donation match that about half of staff use, the company says roughly 20% of its shares will be dedicated to the public good.
What stands out for practitioners is that an AI company is treating job disruption as something to back with equity. Anthropic had earlier committed $200 million to studying and addressing AI's labor market impact. If you plan AI rollouts, adding a column next to the cost-savings figure for which tasks change and where the people doing them move will make the plan easier to explain.
Much is still undecided. Who runs the fund, how grants are decided and for whom, and what valuation serves as the baseline will be shared only in the coming months. Because stockholder approval is pending, even the 5% figure is not final.
herdr — AI coding sessions survive closing the window
herdr is an open-source tool that runs AI coding tools such as Claude Code, Codex, Cursor, and OpenCode in multiple panes of a terminal (the black window where you type commands) and manages them together. It ships as a single Rust binary under the Apache-2.0 license, which is fairly permissive for commercial use. Each pane shows whether its AI is working, stopped waiting for your input, or idle.
The key feature is session persistence. If you close the window or lose a remote connection, the terminals keep running in a separate background server. After a reboot the running processes themselves end, but the README (the guide on the repository's front page) says herdr restores the layout and can resume conversations for supported AI tools. AIs can also open new panes and send each other instructions.
If you have ever lost work when your editor froze while several AI coding tools were running, the benefit is immediate. Because you can see at a glance which AI is waiting for your approval, you can queue several tasks and visit only the stalled panes. On macOS and Linux it installs through Homebrew (a program installer for Mac and Linux).
Windows support is still in beta. The guide does not mention creating a separate copy of the code for each task, so you need to split working folders yourself to keep several AIs from editing the same folder at once.
Lightpanda — an AI browser 9x faster than Chrome
Lightpanda is a headless browser, one that opens and reads web pages in the background without showing a window, built for AI agents and automation. It is not a modified Chrome but a new browser written in the Zig language. In the README's benchmark table, processing 100 web pages used 123MB of peak memory versus 2GB for headless Chrome, about 16x less, and took 5 seconds versus 46, about 9x faster.
The figures come from fetching 933 real web pages over the network on a single general-purpose AWS server. Popular web automation tools such as Puppeteer can connect to it as is, and it can also run as an MCP server, the standard way AI tools plug in outside capabilities. It can export pages as HTML, Markdown, or PDF and has an option to respect a site's crawling rules.
If you run automation that repeatedly reads many pages, such as price comparison, listing monitoring, or competitor page collection, it is a candidate for cutting server cost and wait time, because one server can handle more jobs at once. Start by moving jobs that read text content rather than check how pages look.
The maker ran the benchmark itself, so re-measure on your own target sites. The license is AGPL-3.0, so if you modify the browser and offer it as a service to outside users, you may have to publish your changes. On Windows it runs only through WSL2, a feature that runs Linux inside Windows.
PaddleOCR-VL 1.6 — reads PDF tables and formulas with structure intact
Baidu's PaddlePaddle team released the document parsing model PaddleOCR-VL 1.6 on May 28. It extracts not only text but also tables, formulas, charts, and seals from PDFs and scans while keeping their structure, and at about 0.9B (roughly 900 million settings) it is far lighter than large AI models. The team says it scored 96.33% on OmniDocBench v1.6, a test of document parsing quality.
OCR turns text in images into editable text, but older approaches read only the characters and often scrambled table layouts and formulas. Since its first version last autumn, PaddleOCR-VL has narrowed this gap by understanding images and text together, and version 1.6 is said to improve tables, charts, and rare characters. The model files are on Hugging Face, an AI model sharing site, under the Apache 2.0 license, so you can download and run them on company machines.
If you have fed reports, quotes, or disclosures to AI and got back shifted units or misaligned table cells, it is worth testing this model to extract tables and numbers first, before handing the whole PDF to a general chatbot. Being able to process documents in-house without sending them out also helps where security rules are strict.
The score is self-reported by the developers, and accuracy on Korean forms has not been confirmed in public material. Installation requires a minimum PaddlePaddle version, and Mac users are advised to use Docker (a tool that runs programs as self-contained packages). If this is new to you, setting up a test environment with your IT team will be faster.
Try today
Open your AI rollout plan or automation list and write one line per item: what will the person who did this work do next? Now that an AI company has named job change as something to support, the same question will soon come up in internal reviews.
If you use two or more AI coding tools at once, install herdr on macOS or Linux with brew install herdr, start a task, close the window, and reopen it. Just checking whether the work carries on is enough to decide whether to make it a daily tool.
If you have automation that repeatedly reads web pages, pick one site and process the same batch of pages with Lightpanda and with your current Chrome setup to compare memory and time. If you plan to offer it as an external service, check the AGPL terms with legal first.
Pick one non-sensitive PDF whose numbers often come out wrong, extract its tables with PaddleOCR-VL 1.6, and check numbers and units against the original cell by cell. That one comparison sheet becomes the basis for deciding whether to use the model at work.