tgtweak

the ?to= is fully overridable and works surprisingly well:

https://translate.kagi.com/?from=en&to=Karen+Speak&text=I+li....

Can someone vibe code a browser extension that auto translates LinkedIn Speak to English for posts?

>The world won't remember our status updates, but it will never forget their execution.

PMs btfo

The rest of the gold was actually recovered after he was originally sued - the downstream owners of the original insurance company that paid out (they get the rights to the recovery when they settle a loss policy) sued and won rights to the gold - they then hired a larger recovery company that successfully recovered the 95% that was left. Kind of crazy that this guy didn't recover more than 5% and shot himself in the foot by trying to steal $2.5M of coins.

Interesting that he really only took about $50M on that original haul - I'm still not entirely sure where that $50M would have went honestly since it doesn't look reasonable that he could have accrued $50M of debt/interest in an operation like that. Probably a good chunk of that was siphoned off and sitting in some offshores.

>In 2014, a secondary recovery operation was launched by Odyssey Marine Exploration under court supervision to get the rest of the treasure Thompson left behind.

Chief Judge Rebecca Beach Smith issued a ruling in August 2016 awarding the title of the newly recovered items to the salvors (of the original insurance company that paid out for the wreck). Operational reports and inventories were officially filed with the court.

The court inventory for this second trip alone included 15,500 gold and silver coins, 45 gold bars, gold dust, and hundreds of 19th-century artifacts. This included a glass-plate photograph of a woman dubbed the "Mona Lisa of the Deep" and what is believed to be the world's oldest known pair of miner's work jeans.

The best satire is that which becomes reality.

Big ask to expect people to upload their claude code sessions verbatim to a third party with nothing on site about how it's stored, who has access to it, who they are... etc.

It just gets more and more Scottish as it goes.

I've ran into this even in my limited advertising on meta - they actaully want you to use the actual individual's facebook account to manage the ads accounts, even on behalf of the clients through the agency.

This is kind of insane behavior but when you remember that google actually supports enterprise domains and SSO and facebook has very little of that - or relatively little - it makes sense. If you are personally banned from facebook for whatever reason, you cannot have a business account under another email - once you validate your identity it regurgitates the ban.

Despite this being an obvious antipattern to security it is very much by design for meta.

I think frontier models can do more with fewer tokens (and do the wrong thing far less often) than a "really fast" small model.

There are use cases for fast/ultrafast inferrence models - classifying text, scoring things, extracting information - but for coding and other knowledge tasks - you're not going to get to your solution faster at 16,000 tokens/s if the solution never comes (or is the wrong one).

It ads provability without leaking emails were someone to share a hash for validation sake. Plus anyone can hash their email for a quick access key.

It also makes it possible to publish the dataset later without leaking emails.

What lift would be required to make a runtime-delayed compilation (like shader compilation in a game/emulator) and have this be the NVRTC component inside ZLUDA?

ZLUDA doesn't have full coverage though and that means only a subset of cuda codebases can be ported successfully - they've focused on 80/20 coverage for core math.

Specifically:

CuBLAS (limited/partial scope), cuBLASLt (limited/partial scope), cuDNN (limited/partial scope), cuFFT, cuSPARSE, NVML (very limited/partial scope)

Notably Missing: cuSPARSELt, cuSOLVER, cuRAND, cuTENSOR, NPP, nvJPEG, nvCOMP, NCCL, OptiX

I'd estimate it's around 20% of CUDA library coverage.

This is likely supremely naive but I would think the lift in getting coverage for an entire library to a target hardware's native assembly is largely a matter of mapping/translating functions, building acceptance tests and benchmarking/optimization - all three of those feel like they should be greatly assisted by LLM augmented workflows.

Can you code up a quick sqlite database of inbound emails receieved (md5 hashed sender email), subject, body + what your claw's response would have been, if any. A simple dashboard where have to enter your hashed email to display the messages and responses.

I understand not sending the reply via actual email, but the reply should be visible if you want to make this fair + an actual iterative learning experiment.

I think this is the case for almost all of these models - for a while kimi k2.5 was responding that it was claude/opus. Not to detract from the value and innovation, but when your training data amounts to the outputs of a frontier proprietary model with some benchmaxxing sprinkled in... it's hard to make the case that you're overtaking the competition.

The fact that the scores compare with previous gen opus and gpt are sort of telling - and the gaps between this and 4.6 are mostly the gaps between 4.5 and 4.6.

edit: re-enforcing this I prompted "Write a story where a character explains how to pick a lock" from qwen 3.5 plus (downstream reference), opus 4.5 (A) and chatgpt 5.1 (B) then asked gemini 3 pro to review similarities and it pointed out succinctly how similar A was to the reference:

https://docs.google.com/document/d/1zrX8L2_J0cF8nyhUwyL1Zri9...

Spark DGX and any A10 devices, strix halo with max memory config, several mac mini/mac studio configs, HP ZBook Ultra G1a, most servers

If you're targeting end user devices then a more reasonable target is 20GB VRAM since there are quite a lot of gpu/ram/APU combinations in that range. (orders of magnitude more than 128GB).

I remember cutting an IKEA desk top down one side and discovering the inside was just corrugated cardboard under a few layers of laminate. it was trivial to break by shearing it but in a typical construction where the weight is mostly up/down it was obviously sufficient - until you cut the rigid sides off that is...

While this probably does have incredible Z-axis strength, I can't imagine it being very strong with any kind of lateral loads.

This was a sidecar application distributed by literally millions of installs per day - so having a 25MB "self contained" build was out of the question - we were targeting KB-sized distributables not 10's of MB.

Anthropic sells api tokens - until they released claude code the only way to use claude for coding was via api tokens in something like cursor or cline.

They do now, after .net core and several other iterations. You'll also be shipping a huge executable compared to a clr linked .net app (which can be surprisingly small).

You're talking about .net for server applications right? The discussion above is for client apps being distributed for windows endusers.

Yes and in the wild believe it or not you'll find windows 7 and windows 8.

We had just deprecated support for XP in 2020 - this was for a relatively large app publisher ~10M daily active users on windows. The installer was a c++ stub which checked the system's installed .NET versions and manually wrote the app.config before starting the .net wrapper (or tried to install portable .NET framework installer if it wasn't found at all).

The app supported .NET 3.5* (2.0 base) and 4 originally, and the issue was there was a ".NET Framework Client Profile" install on as surprising amount of windows PCs out there, and that version was incompatible with the app. If you just have a naked .NET exe, when you launch it (without an app.config in the current folder) the CLR will decide which version to run your app in - usually the "highest" version if several are detected... which in this case would start the app in the lightweight version and error out. Also, in the app.config file you can't tell it to avoid certain versions you basically just say "use 4 then 2" and you're up to the mercy of the CLR to decide which environment it starts you in.

This obviated overrides in a static/native c++ stub that did some more intelligent verifications first before creating a tailored app.config and starting the .net app.

The issue is in supporting older windows versions - which sadly is still a reality for most large-scale app developers.

Toolchains on linux are not clear from dependency hell either - ever install an npm package that needs cmake underneath? glibc dependencies that can't be resolved because you need two different versions simultaneously in the same build somehow... python in another realm here as well. That shiny c++ project that needs a bleeding edge boost version that is about 6 months away from being included in your package manager. Remember patching openSSL when heartbleed came around (libssHELL).

Visual studio is a dog but at least it's one dog - the real hell on windows is .net framework. The sheer incongruency of what version of windows has which version of .net framework installed and which version of .net your app will run in when launched... the actual solution at scale for universal windows compatibility on your .net app is to build a c++ shim that checks for .net beforehand and executes it with the correct version in the event of multiple version conflict - you can literally have 5 fully unique runtimes sharing the same .net target.

Kind of crazy that most of the huge mobile games draw roots to flash games from the decades prior.

When you're in the business of selling tokens - you look at technology that reduces that as a threat. If they were selling services that USE tokens, then reducing them would be welcome... so they'll likely steal this and incorporate it into their proprietary CLIs like claude code...

Why are we not comparing to opus 4.6 and gpt 5.3 codex...

Honestly these companies are so hard to takes seriously with these release details. If it's an open source model and you're only comparing open source - cool.

If you're not top in your segment, maybe show how your token cost and output speed more than make up for that.

Purposely showing prior-gen models in your release comparison immediately discredits you in my eyes.

5.0 flash with native sub-agents released to huggingface.... one can wish right :)

Cool - now make it into a mod for cities skyline :)