4 things that happened while you were busy
1 Anthropic's IPO filing reveals $16B in revenue and an $80B loss
According to the Wall Street Journal, which reviewed the filing, Anthropic brought in $16B in the first half of the year and posted an $80B net loss, mostly because of accounting changes. It could spend as much as $750B on computing power, and the listing is expected in mid-November. Meanwhile, OpenAI is reportedly racing to close a $30B funding round.
2 Perplexity Computer can now run on your Mac and keep your files private
Perplexity's agent can now work locally on a Mac with at least 36GB of RAM. It uses a hybrid approach: models on your computer handle your files, while cloud models do the heavy reasoning. It follows Meta's open model, which was also tuned for Macs last week.
3 ElevenLabs and Krea launch new voice and image models
ElevenLabs released Eleven Realtime, a speech-to-speech model for voice agents that respond as you talk. Krea launched Krea 3, an image model built for editing at scale, which ranks fourth on Arena's multi-image editing leaderboard, behind only OpenAI.
4 Two new models chase speed and low prices
Cloudflare's Clairic pulls specific values out of messy text in a single pass and returns nothing when a value is missing instead of guessing. Cloudflare says it is 3.2 times faster and 49% cheaper than using a regular LLM. Inception's Mercury 3 generates more than 1,000 tokens a second with a 1M-token context window, starting at $0.10 per million input tokens, and its makers claim it beats rival small models on coding.
The meme. A popular joke shows someone simply telling an AI to fix a hard problem. In Silicon Valley, some people are even smoking again because they believe AI will cure cancer anyway. Citrini Research, a five-person financial research team with no scientists, decided to test the idea for real by asking Claude for a new scorpion antivenom.
What happened. Over three months, Claude generated 211 possible antivenoms, and 17 met Citrini's criteria. Claude did nearly all of the work. It took about 367 messages, 10,000 hours of H100 GPU time, early access to Anthropic's internal models, and a $200,000 budget.
The catch. The candidates were shown to attach to the scorpion toxin, but nobody tested whether they actually neutralize it. There has been no animal testing, no dosing work, and no safety work yet.
Why it matters. A small team with no lab got from an idea to promising candidates in a few months. But finding a candidate is the cheap part of drug discovery. The open question is whether AI can help with the slow, expensive testing that comes next.
A cheaper AI model matched a frontier one once it had guardrails
The test. Naresh Jain, founder of Specmatic, wanted to know whether a cheap model could be made reliable with a "harness": written specs, examples, and tests that flag mistakes while the model is coding. He had GPT-6 Sol and the cheaper GPT-6 Luna build the same Java backend 10 times each, with and without the harness, then checked every build against 178 hidden tests. The project brief is public.
The result. On its own, Luna passed only 3 of 10 runs. With the harness, it passed all 10, matching Sol at $0.273 per run versus $0.406. Adding the harness to Sol raised its cost about 7 times with no gain in quality.
What to take from it. Keep the harness on when you use cheaper models. For frontier models, run the tests once at the end as a final check. And keep your specs up to date with human review, since the harness is only as good as what it checks. Jain calls this "executable attention": failing tests that catch mistakes early, before they pile up.
📈 Get Recommended by AI. Rob Hoffman shares 11 website changes, such as clear labels, plain text, and well-structured HTML, that make it easier for AI agents to recommend your product (8.5K bookmarks).
🔒 Lock Your Chats. ChatGPT now offers Locked Chats, a PIN-protected space for sensitive conversations, with Face ID or fingerprint unlock on supported phones.
💨 Speed Up Your Computer. A viral thread offers seven prompts for using the Claude desktop app to diagnose a slow computer, clear startup items, free up storage, and set up backups (25K+ bookmarks).
📉 Freelance Design Under Pressure. A Redditor who lost a top client to ChatGPT asked if AI has ended freelance web design. The replies include a refreshing counterpoint.
😆 Five Things I Always Say. A TikTok trend has people asking chatbots to list five things they always say, and the answers range from sweet to brutal.
🧑💻 Advice for Juniors. Amp's Thorsten Ball shared his reply to junior engineers who ask what to do next, including joining a company that lets you use as many AI tokens as possible (3.7K likes).
🖥 A Supercomputer on Your Desk? Charles Hoskinson argues that six of the coming Mac Studio M5 Ultras could run a 2-trillion-parameter model for about $100K, far less than comparable Nvidia hardware (1.4M views).
