meatmanek

There are 1-bit* LLMs (+1/-1, usually), though they often come with an additional scaling parameter:

https://huggingface.co/prism-ml/Bonsai-27B-gguf#weight-repre...

Weight Representation: Q1_0_g128 Each weight is a single sign bit: 0 maps to −scale, 1 maps to +scale. Every group of 128 weights shares one FP16 scale factor. Effective bits per weight: 1.125 (1 sign bit + 16-bit scale amortized over 128 weights)

The Bonsai ternary model also uses this group-wise scaling, so the theoretical bits per parameter is 1.71 instead of ~1.58; additionally, they actually store trits as 2 bits; with the group scaling factor it becomes 2.125 bits/weight: https://huggingface.co/prism-ml/Ternary-Bonsai-27B-gguf#memo...

Scan the backs

I guess this is assuming some scanning rig (camera scanning?) where it only takes a few seconds per photo. With a flatbed scanner, scanning the backs is almost certainly going to take longer than just typing dates in manually.

My (probably mis-)understanding of StubHub/Ticketmaster is that they're serving as sin eaters / reputation shields for the artists:

Phase 1:

   - Artist lists tickets as "$20"
   - SH/TM charges $20 + $15 convenience fee
   - SH/TM pays the artist $30 and keeps $5
   - Artist gets to look like they're kind/generous for pricing so affordably while making some of SH/TM's cut

   - SH/TM hold back some percentage of tickets (possibly most or all of them) to sell on the "resale" market
   - Those "resale" tickets sell near the market-clearing price, either earning higher margins than if they had sold at list price, or ensuring that most of the seats in the venue are filled.
   - SH/TM kick back some percentage of revenue from resale tickets to the artist.

It's also super fast on CPU.

Direction dependence also means your output drops even more on cloudy/overcast days.

It would be really cool if they could expose this information to customers somehow. Imagine:

   - having a log of the most prominent J-space tokens during your customer support chatbot's interactions with a user, so you can have more introspection into why a particular outcome happened
   - being able to detect certain thoughts associated with undesirable behavior (hallucinations, overstepping authority, lying, etc.) and trigger some sort of remediation (e.g. upgrading to a better model, redirecting to a human, forcing tool calls)

All you're saying is that the government under Biden talked to Meta.

https://youtu.be/zdA__2tKoIU?t=149

You are missing the context of the ecosystem

That's one of several reasons why a citation is warranted.

This is super cool. Do you know if any of the inference backends (llama.cpp, vllm, etc) support this technique?

speed dial options. That doesn’t happen if you don’t click on them directly.

It also seems to happen if you type the domain name in the address bar but hit enter when the suggested URL autofills. For me, typing out aliexpress.com fully will send me directly to AE, but typing aliexpress.c and hitting enter (with the autofill completing "om") redirects through vivaldi.com/bk/aliexpresscom-us

Frankly, it's not really more insecure than any other installation method. Apt packages and the like generally have the ability to specify pre/post-install scripts, so `sudo dpkg -i ./random.deb` is equivalent to `sudo bash ./random.sh`. Even if they didn't have pre/post-install scripts, they're still writing arbitrary files to arbitrary locations on your disk, so they can trigger execution the next time you boot or log in or whatever.

And at the end of the day, no matter the installation method (even just unpacking a tarball and executing the program directly from that directory), you're going to run their program on your computer, and then the program can do whatever it wants. Maybe you don't run it with sudo, but https://xkcd.com/1200/ seems relevant.

Freemium is a potentially interesting model for app developers. If you're small enough to be in the free tier, you can avoid the risk of someone turning your app into the next backend for Chipotlai Max and running up your AI bill.

Once you get enough downloads to where Apple starts wanting to charge you money, you can consider switching to OpenAI or Anthropic or Google or Deepseek or whatever. Sounds like they've even made that relatively easy to do in the Foundation Models framework -- just implement the LanguageModel protocol. I expect open-source or provider-written adapters to pop up that would let you use your vendor of choice.

I'm guessing they'll integrate with the double-tap-the-bottom-of-the-screen feature that pulls up siri in front of a screenshot. Currently it doesn't seem to hook into "visual intelligence", and needs to call out to ChatGPT to do anything with the screen contents.

"shape"

At least that somewhat aligns incentives between players and the game studio. If an old game has a long-lasting player base, then a modest subscription makes it more likely that the studio would keep the servers up and running, if not actively patching the game. With a game that you pay for up-front, a long-lived player base can be a liability for the company (ongoing costs without many new purchases.)

it's a joke: the author of the post is literally the guy who wrote PuTTY

I still use google as a verb even though I use Kagi.

Presumably asymmetric splits meaning A and B can talk, B and C can talk, but A and C can't talk.

AI works great on it.

Define AI, and define "great"

What they first received

Shouldn't every cell in this column be the same?

I think you're correct with the standard thinking approach (just generate a big stream of tokens before drafting your actual answer). After a while, additional thinking just results in loops.

The RSA approach from https://rsa-llm.github.io/, expanded on by https://www.zyphra.com/post/zaya1-8b, looks like a promising way to squeeze a bit more intelligence from a small model. As I understand it, running multiple independent thinking traces in parallel gives you a chance of one of them finding a different local optimum, whereas running a single trace for longer is likely to just circle around one optimum.

That said, at the end of the day, there's only so much information a small model can contain. If a model just doesn't know some key piece of information, no amount of thinking will help it figure out a solution that depends on that information.

Doubleclicking random exe files off the internet is almost uniquely a Windows problem.

Tell that to my partner's grandfather, who managed to find and install malware chrome extensions on his chromebox.

Ok, I double-checked, and I get 21-22tps with lmstudio-community/LFM2-24B-A2B-Q4_K_M.gguf running under LM Studio on my i5-12400 with 2x32GB sticks of DDR4 3200. This is with small context (just "Write me a poem about a language model named Liquid" in `lms chat`)

    Prediction Stats:
      Stop Reason: eosFound
      Tokens/Second: 21.10
      Time to First Token: 1.827s
      Prompt Tokens: 42
      Predicted Tokens: 187
      Total Tokens: 229

This model is pretty cool if you don't have a GPU - I was able to get I think 20 or 30 tokens per second on CPU (DDR4 ram) alone. (I don't remember if that was with q4 or q8.)

Otherwise, if you have a GPU with more than like 4GB of VRAM, there are better models. Gemma4 and Qwen3.6 (or Qwen3.5 if you need the smaller dense models that haven't yet been released for 3.6) are a good place to start.

It's not that surprising that an 8B dense model would compete with a 35B-A3B MoE model.

The geometric mean rule of thumb for MoE models is that the intelligence level of an MoE model with T total parameters and A active parameters is roughly equivalent to that of a dense model with sqrt(A*T) parameters. For qwen3.6-35B-A3B, that equivalent size is 10.24B, spitting distance of an 8B model. Good training can make up the 28% difference in size.

Several years back when I applied for a Google internship, I missed some emails from my recruiter (soandso@google.com) because they went to my gmail spam folder.

The Unsloth llama.cpp guide[1] recommends building the latest llama.cpp from source, so it's possible we need to wait for LM Studio to ship an update to its bundled llama.cpp. Fairly common with new models.

1. https://unsloth.ai/docs/models/gemma-4#llama.cpp-guide

By the same logic, wouldn't 4 months of severance pay be equivalent to forfeiting 92% of salary?

For something paid at regular intervals like RSUs, you really should never be looking at the total value of the grant, and instead think of it in terms of how many shares per paycheck/month/quarter/year you vest.

If you've got a cliff coming up, that's different. I'd be pissed if a company laid me off 11.5 months into a 12 month cliff or a few weeks before an annual bonus and didn't accelerate the vesting / bonus.

Be careful working CTRL + W into muscle memory though, I've lost count of how many browser tabs I've closed by accident...

I still maintain this is why macOS is the best OS for terminal work -- all the common keybindings for GUI tools use a different modifier key, so e.g. ⌘C and ⌘W work the same in your terminal as they do in your browser.

(Lots of the readline/emacs-style editing keybindings work everywhere in macos as well -- ^A, ^E, ^K, ^Y, but not ^U for some reason)