Worth taking a closer look at what they actually were able to formally verify. No doubt a massive achievement, however after seeing several demos poke holes through it I don’t think they’ve done the level of “formal verification“ that you may be inclined to think they have.

One of the main goals at the moment is to keep thinking human legible. You can imagine how much harder it’d be to do root cause analysis on the recent OpenAI event if we couldn’t even tell what they’re thinking.

Claude design's prompt is trivial to verify. They bundle it in the frontend bundle and send it on every network request.

CoT legibility largely disappears which is quite concerning from a safety perspective

Just saying you’re not alone, very surprised by the reception given how brutally sloppified the OP is.

Interesting problems space but I hope the author just gives dot points next time rather than bloating it and losing most of its meaning.

Could you please elaborate a bit more for my understanding?

What in particular about this method breaks correct token boundaries?

On my first read I read your comment as there are special tokens that require multiple tokens to emit, hence you can't get certain tokens emitted alone - but I don't think that's what you're getting at on a second read?

Interesting that you've found similarities between "d" and the hidden tokens for opening an xml tag, pressing caps lock and the other hidden tokens of note. I haven't run into any trouble extracting "d" tokens, is it a particular model that you see create that pattern?

That's " 'd ".strip(), an english contraction suffix. it's 1 token, but using this echo approach you will be served the apostrophe and the subsequent letter for the first time in different steps.

I couldn't reproduce this behavior with Sonnet 4, and Sonnet 3.7 has been deprecated since I messed with this stuff. You can try tokenizing the string "<hello> </hello>"

I think the correct tokenization of the string will not have any tokens that contain mixed punctuation and letters, but the result of this approach does contain such claimed tokens.

I use bedrock with 1M context every day. Not sure this is right

4.7 is the first opus model that’s had the 1 M context window available on Bedrock.

Not true. Opus and Sonnet 4.6 support 1m context on Bedrock.

I've had Opus 4.6 1M and Sonnet 4.6 1M for months now on Bedrock.

Their docs may be lying but they say 200k for opus 4.6. And yes 1M was on sonnet for Claude enterprise.

A lot of enterprises were doing that but now they hit the 150 user limit on Claude and are paying seat+api rates.

Codex is still going strong but it’s hard to imagine they won’t do similar eventually.

So now im honestly hearing a lot more folk stick it out with cursor while waiting for the dust to settle.

>A lot of enterprises were doing that but now they hit the 150 user limit on Claude and are paying seat+api rates.

A lot of enterprises use Github Copilot which has per-request pricing model which effectively means unlimited tokens which eliminates this issue.

Github Copilot changed to token pricing earlier this week

Oh oh :(

This isn't true anymore.. GitHub are rate limiting people on pro and pro+ pretty hard

I'm talking about per-request model remember? With extensive prompt you realistically can have one request every 10 minutes because the agent will be busy for at least 10 minutes executing it. They aren't rate limiting that.

Each tool call requires a new request no?

The harness receives a response, has to parse out the tool call, execute it and then start a new request with the tool call result.

>Each tool call requires a new request no?

Nope, not unless you are doing steering.

Each new prompt = new request, but tool calls don't count.

There’s a lot of tradeoffs to play with, those inference ASICs may not carry the gradient but they are still optimised for larger batches and to run any model. They need enough memory for the weights, wide batch inference, and ideally leftovers for kv cache efficiency.

For personal inference you’re given a lot more room to play in - much of it poorly explored today - enough to concern an argument of cost advantages evaporating

You can ablate surprisingly large chunks of a model with near to no effect, you can try this easily - download an open weight model in torch.

Obviously it’s not ideal but you could likely have single digit % of all weights affected and still have a useful model (many caveats here: e.g. locality of damaged weights matters, distribution of errors matters, fail high/low matters, …)