I built The Timeline because I was tired of asking "where does this stand?" and trying to remember what me and my agents even did yesterday.
It joins Slack and Telegram, pulls in GitHub, Linear, Sentry and more, and keeps one work record that updates itself. Ask anything, get answers with sources. It also builds and maintains your boards and lists (tasks, people, decisions…).
Agents are teammates here: they read the same context and report their work back to it.
It's early and free. Looking for a few teams to try it and tell me what's broken. Design partners? 👉👈
https://t.co/nFIa9xibT4
Most queues treat retries as free. With LLM calls they're not.
A model step that times out after the provider already billed you, then gets retried, is the same tokens paid twice. Record the spend first, then decide if it's worth retrying.
Shameful…and naive. A senior American official criticizing Ukraine, a democratic European country, for defending itself against Russia. Not quite how Ronald Reagan would have done it.
Anthropic goes public this week, so I read the part of its prospectus about open-weight models.
In the risk factors, Anthropic writes that open-weight models "may capture a portion of the market for certain use cases, which could exert pricing pressure on our offerings".
Databricks published data on Monday that shows what that looks like. On models Databricks hosts, open-weight models went from 7.3% of tokens in July to 30.9% in early November. Companies using just one model provider fell from about 64% to 28%. Databricks sells the router, so they like this story, but these numbers come from usage on their gateway.
The same direction shows up elsewhere: open-weight models went from 26.8% to 71.7% of tokens among OpenRouter's top 50 models over the past year, and Menlo Ventures estimates their share of enterprise tokens rose from 11% to 19%.
The same Databricks data shows Anthropic doing fine. Claude Opus 5.5 went from 7.2% to 37.9% of proprietary model tokens in six weeks.
So people pay for the best model when the task needs it, and move the rest to whatever finishes the job for less.
That's how I want to work too. Pick the model per task, use my own key, switch when something better or cheaper ships. That only works because open-weight labs publish their weights and other companies can host them.
As a customer, that's the risk factor I want to see come true.
https://t.co/FVKdu2jeaO
Dario WON! This is possibly worst thing that can happen to Open Models.
1. Data centres are cooked. In SE-Asia that rent capacity to Chinese customers have been a channel for NVDA demand and this rule puts that channel under licensing.
2. Seven more countries need licences (Singapore, Malaysia, Thailand, Indonesia, Brazil, South Africa, and Norway). Buyers there must disclose the identity of the end users and the intended use of the hardware to ensure the chips are not rented to Chinese firms to bypass U.S. export bans.
3. Now for Chinese AI labs getting top Nvidia chips for training will be more harder + slower. Some Chinese labs have contacted overseas data-center operators about reserving NVDA chip capacity but now they will face more pressure to rely on domestic chips / existing stockpiles.
It is so over.