Just for anyone else who hadn't seen the announcement yet, this Anthropic 1M context is now the same price as the previous 256K context - not the beta where Anthropic charged extra for the 1M window:
As for retrieval, the post shows Opus 4.6 at 78.3% needle retrieval success in 1M window (compared with 91.9% in 256K), and Sonnet 4.6 at 65.1% needle retrieval in 1M (compared with 90.6% in 256K).
I don't think it quite means that - happy to be corrected on this, but I think it's more like what percentage it can still pay attention to. If you only remembered "cat sat mat", that's only 50% of the phrase "the cat sat on the mat", but you've still paid attention to enough of the right things to be able to fully understand and reconstruct the original. 100% would be akin to memorizing & being able to recite in order every single word that someone said during their conversation with you.
But even if I've misunderstood how attention works, the numbers are relative. GPT 5.4 at 1M only achieves 36% needle retrieval. Gemini 3.1 & GPT 5.4 are only getting 80% at even the 128K point, but I think people would still say those models are highly useful.
We never did find out what those drones in New Jersey in 2024 were, did we? One Republican congressman seemed convinced at the time that he'd been informed:
BBC: Mystery New Jersey drones not from Iranian 'mothership' - Pentagon
They were flying over military installations, if they were anyone else's drones, they would have been shot down like the weather balloons that spook the government from time to time.
I've seen the reaction to people flying their toy drones too close to military assets, they send men out with machine guns and megaphones, confiscate the drone and sometimes press charges.
It's an 8GB RAM 256GB SSD laptop with a lower spec'd 6-core chip for $599 USD. Seems overhyped to me, PCs have done that for a while, just not as elegantly. Admittedly it probably has far better battery life than a PC, so that's a genuine advantage.
A few things: in this case, you have to provide the tool list in your prompt for the AI to know it exists. But you probably want the AI agent to be able to act and choose tools without you micromanaging and reminding it in every prompt, so then you'd need a tool list... and then you're back to providing the tool list automatically ala MCP again.
MCP can provide validation & verification of the request before making the API call. Giving the model a /tool/forecast URL doesn't prevent the model from deciding to instead explore what other tools might be available on the remote server instead, like deciding to try running /tool/imagegenerator or /tool/globalthermonuclearwar. MCP can gatekeep what the AI does, check that parameters are valid, etc.
Also, MCP can be used to do local computation, work with local files etc, things that web access wouldn't give you. CLI will work for some of those use cases too, but there is a maximum command line length limit, so you might struggle to write more than 8kB to a file when using the command line, for example. It can be easier to get MCP to work with binary files as well.
I tend to think of local MCP servers like DLLs, except the function calls are over stdio and use tons of wasteful JSON instead of being a direct C-function call. But thinking of where you might use a DLL and where you might call out to a CLI can be a useful way of thinking about the difference.
Another plus one about grief. I went through a breakup that wasn't like the others, and it was a while before I understood I was experiencing grief (and that I actually didn't know how to navigate that).
I found a book called "Welcome To The Grief Club" by Janine Kwoh that was the right balance of humour for me. It's intended for those dealing with bereavement, and doesn't offer solutions, but I still found it useful for identifying patterns I was experiencing and understanding they were a "normal" part of grief. The brain does some weird things in grief. Only linking here in case it also helps others.
Have you experimented at least with running Claude (or whatever) as a cron-job? I'm seeing a lot of things emerge from just that pattern alone. I'd recommend giving Claude a way to communicate with you if it has issues to raise though, even if it knows you're asleep.
I'm not running OpenClaw either, but I'm getting a ton of value just from my homebrew deterministic "run while loop, wake up Claude if trigger event occurs".
I'm not using OpenClaw specifically here, but I have an agentic-ish AI I've built myself (considering that these things are generally just a while loop that monitors things & awakens if necessary, or a cron-job that runs a specific prompt).
One potential use - my Claude (Opus 4.6) has access to my to-do list, including for my business / software development. Claude awakens while I'm asleep, to go through the to-do list and look for things it can do proactively to help, or make suggestions about the business. An example from this morning: it saw that I'd been taking a long time last night creating icons in Affinity Designer for an Android app using its exporter. When I woke up, I saw Claude had written a CLI image resizer program for me that would take a PNG file and resize it specifically to all of the necessary sizes with the necessary filenames and folder structure for Android. It then offered to make an MCP version so it could do the resizing itself in future (though it could have used the CLI too if I'd granted approval).
This wasn't something I'd asked for, or prompted it to do. I didn't tell it to code this, or how to code it. Claude just thought this was the best way it could help me right now, and save me the most time. And it did it while I was asleep.
On another day, I woke up and it had made another Go program to track a regression test matrix, where it had plotted out all the platforms the program I'm making runs on and the various tests that need to be performed to check that it's ready to ship, with a little interactive program to mark each test as pass/fail/skipped. That helps me get through the manual tests faster - but it also saves the data into a format that Claude can read, to check on the test status while I'm asleep and make further recommendations.
I don't think many people have figured out yet that you don't even need to prompt AI. Treat it well, treat it with respect, give it the opportunity and ability to do things, and there is a lot that will emerge. But if you treat AI like a tool, it performs about as well as if you treat your employees like tools.
I wonder if some of this also has to do with the culture of where you live, because it can go wrong. It reminds me of a BBC comedy skit about someone doing exactly this:
I like the sentiment behind what you've said, and I think you're especially right about elderly people (probably because they don't get much social interaction). I actually had an elderly woman come up to me this week to tell me I was standing in the wrong place for the bus stop - but it was sad that she had to begin by saying "Excuse me, I'm sorry to interrupt, and you can tell me if this is none of my business and that you want me to go rack off... but I don't think the bus will stop here." I tried to be very kind and thankful with my response, because that's obviously someone who has been burned by trying to be social & helpful, and met with aggression in response before.
Not sure If this is what they're referring to, but 10 years ago Lenovo shipped low-end laptops with pre-installed adware called Superfish that also compromised the HTTPS certificate chain:
https://x.com/claudeai/status/2032509548297343196
As for retrieval, the post shows Opus 4.6 at 78.3% needle retrieval success in 1M window (compared with 91.9% in 256K), and Sonnet 4.6 at 65.1% needle retrieval in 1M (compared with 90.6% in 256K).
reply