1,798 karma · joined August 12, 2011
As for pricing, that model's pages says: "Predictions run on Nvidia A100 (80GB) GPU hardware. Predictions typically complete within 17 seconds."
And the pricing page (https://replicate.com/pricing) says Nvidia A100 (80GB) GPU hardware costs $0.0032 per second.
So Llama 2 70B would "typically" cost under 17 x 0.0032 = $0.0544 per run.
"Stop Energy is not reasoned, it never takes into account the big picture, it is the mirror image of Forward Motion. In the Stop Energy model, everyone, no matter how small their stake in a technology, has the power to veto. Nothing ever gets done, and people who want to move forward are frustrated in every attempt to move. Unfortunately, Stop Energy is the rule, not the exception"
https://radio-weblogs.com/0107584/stories/2002/05/05/stopEne...
I'd make a fast approx check for some of these properties that can run during brute force, then feed anything that passes that check to a better check, then finally to a human (maybe nowadays put an LLM before the human stage).
There's a chance to miss the translation if it doesn't match other messages but this seems like a decent approach.
Having read about JavaScript/Java VM optimisations in JITs and GC I would be surprised if a global state change like this is not manageable - think deoptimising the JS when you enter the debugger in the dev tools in your browser.
The answer is, you model those as well and work out what to do. But it's more messy than you might think if you just model the first-order failure paths.
https://community.openai.com/t/bug-incorrect-chatgpt-chat-se...
I only reported it on the forums because there didn't seem to be an official bug reporting channel, just a heavyweight security reporting process.
As well as the actions they took to fix this specific bug, another useful action would be to have a documented and monitored bug reporting channel.
I reported this bug in the forums (no response), because I couldn't find an official way to report bugs - at least it wasn't documented anywhere that I could find. The actions they've taken fix that bug, but not the next one. OpenAI should take an action to improve their bug reporting channels, so the next bug gets found more quickly.
https://community.openai.com/t/bug-incorrect-chatgpt-chat-se...
Other messages indicate issues with bug reporting too:
* https://news.ycombinator.com/item?id=35291943 * https://news.ycombinator.com/item?id=35295747
https://github.com/webassembly/threads
"This proposal adds a new shared linear memory type and some new operations for atomic memory access. The responsibility of creating and joining threads is deferred to the embedder."
As for the current status, it is listed on the proposals page: https://github.com/WebAssembly/proposals
The "Threads" proposal is in "Phase 3 - Implementation Phase" which means it's starting to be implemented, but it hasn't reached "Phase 4 - Standardize the Feature" where 2+ browsers implement it. The final stage would be "Phase 5 - The Feature is Standardized. See https://github.com/WebAssembly/meetings/blob/main/process/ph...
It appears that someone was able to link an MS account to my email with no verification, then rename the account, again with no verification.
Best case is that it's someone who used my email as a recovery email for their MS account and changed it. But with the mess of MS accounts, I'm always nervous they've got some residual control of my real account which is also linked to that email.
Unfortunately I've never been able to get confirmation from MS that things are OK. There are plenty of questions about this particular renaming issue on the web, but no answers.
I get that some people prefer it but a lot of people don't and it's frustrating to be stuck with it. My thinking is that unless there's clearly one best way, and if it's something people have strong opinions about, then a switch or knob to adjust is a good idea.
But I still think we would want a curated collection of chat/assistant training data if we want to use that language model and train it for a chat/assistant application.
So this is a two-phase project, the first phase being training a large model (GPT), the second being using Reinforcement Learning from Human Feedback (RLHF) to train a chat application (InstructGPT/ChatGPT).
There are definitely already people working on the first part, so it's useful to have a project focusing on the second.
"a refinement type is a type endowed with a predicate which is assumed to hold for any element of the refined type. Refinement types can express preconditions when used as function arguments or postconditions when used as return types"
- https://en.wikipedia.org/wiki/Refinement_type
There's even some literature linking contracts and types directly:
"Traditional static type systems are effective for verifying basic interface specifications. Dynamically-checked contracts support more precise specifications, but these are not checked until run time, resulting in incomplete detection of defects. ... This paper explores the key ideas and implications of hybrid type checking, in the context of the λ-calculus extended with contract types, i.e., with dependent function types and with arbitrary refinements of base types."
https://twitter.com/spokesbird
It wasn't quite as exciting as when Stephen Fry's film crew met him. The video of that encounter is the origin of the Party Parrot emote which you might have seen on Slack.
Well, it turns out that it's based on the fact that A0 is defined as having an area of ~1m², then each subsequent size halves the area. So A4 has dimensions that give an aspect ratio √2 and an area of ~1/16m².
A0: 1,189 mm × 841 mm = 999,949mm² ≈ 1m²
...
A4: 297 mm × 210 mm = 62,370mm² ≈ 1/16m²
"The A0 base size is defined as having an area of 1 m² ... rounded to the nearest millimetre, 1,189 mm × 841 mm".
https://www.rfc-editor.org/rfc/rfc8615#section-3
Registrations MAY also contain additional information, such as the
syntax of additional path components, query strings, and/or fragment
identifiers to be appended to the well-known URI, or protocol-
specific details (e.g., HTTP [RFC7231] method handling).
IndexNow would need to change the semantics of how they handle directories, as a key authorises only subdirectories.
I also notice there is an option for changing the filename of the IndexNow key file, but there is less flexibility about the directory it's hosted:
https://<searchengine>/indexnow?url=http://www.example.com/product.html&key=af4c4e043c7d42afad6bdeeda948527d&keyLocation=http://www.example.com/myIndexNowKey63638.txt
Another interesting feature I saw in the standard is that you can host keys in subdirectories too.
"the location of a key file determines the set of URLs that can be included with this key. A key file located at http://example.com/catalog/key12457EDd.txt can include any URLs starting with http://example.com/catalog/ but cannot include URLs starting with http://example.com/help/."
The idea is that you can push a notification to a search engine when content changes instead of waiting for the crawler to notice.
https://<searchengine>/indexnow?url=url-changed&key=your-key
You can also submit more than one URL with a POST.
You can notify Bing at https://www.bing.com/indexnow?url=url-changed&key=your-key
If you notify the IndexNow API endpoint it notifies Bing plus other search engines on your behalf:
https://api.indexnow.org/indexnow?url=url-changed&key=your-k...
This announcement is about how CloudFlare can now do this automatically for sites it hosts.
Some other hosts and CDNs support IndexNow, eg Akamai. See: https://blogs.bing.com/webmaster/october-2021/IndexNow-Insta...
Nudges don't need to steer the user to a specific choice, just a behaviour change. Sticking with a conversion flow counts as a behaviour change.
Nudges don't need to be simple or understandable. They can be a set of complex changes where causation isn't clear. They just need to get results.
The only really hard requirement that would rule out a nudge is if you forced a choice or used financial incentives.
If you read the Nudge book you'll see that it's a political book, really. The authors introduce nudges as an alternative to hard regulation. Instead they propose that governments consider influencing behaviour in a softer way, but still leave the escape hatch open for people with strong preferences to choose what they want. This strikes a balance between state involvement and principles of liberty. (Or at least that's their argument.)
Because of this framing a nudge is defined mostly by what it isn't. It's not a nudge if it forces a user to a choice; a nudge is anything you do that changes what users do without forcing them.
This is what you've done with your series of changes that resulted in increased conversion. You've left all the choices open still, so users have as much freedom as before, but you've managed to predictably change user behaviour in a way that aligns with your goals. In other words, you've nudged them.