karmakaze

I don't think it qualifies for this part:

> The communication must be made in a manner in which another person may view it.

Even 'transmitted' is too broad if you also consider iCloud backup to be a means.

The high volume of posts (and corresponding lower signal quality) cured me. Now I visit see a handful of stories that I didn't need to read. Occasionally worthwhile so still visit but not frequently like I used to.

Can Gemini Flash Lite 2.5 be made to return raw search results. Some 'search' providers I looked at returned summaries, or vector relevance matches (of presumably a smaller/stale page set).

I was just looking into these as DeepSeek Harness w/ Qwen3.8-27B relies heavily on search. I was going to go with Serper.dev[0] $1 per 1000 (or lower in quantity).

The providers[1] behind this Web Search API have very different rates:

    Ceramic.ai: $0.25 per 1,000 requests
    Linkup:     $5.00 per 1,000 requests
    Exa:        $7.00 per 1,000 requests

[0] https://serper.dev/

[1] https://developers.cloudflare.com/web-search/providers/

A $5000 GPU with a TDP of 575W isn't the kind of thing that goes in a typical "Gaming PC". I love that it exists though. Can't wait for AMD to release a competitor and get some sane pricing (if AI demand for wafers levels off).

It says make it your own. Then

> Gadgets you build with the Muse Gadget SDK need a token. Add it to your SDK configuration so your gadgets can pair with the Muse app.

> Each token can be used by a limited number of devices.

> The Muse Gadget SDK is intended for personal tinkering and is not a supported product or developer platform. The SDK can change, break, or stop functioning without warning.

Doesn't really feel like it's my own.

The first (quoted) tweet says "Is Al really going to kill us all? The scenarios are preposterous, and the presumption of inevitability encourages fatalism, panic, and distraction from the more mundane and realistic safety challenges."

This in itself I find problematic. Solving the mundane and realistic safety challenges are what prevents the potential (preposterous) outcomes. There's no need to ignore anything other than the sensationalism like anthropomorphising. The Paperclip Maximizer is meant to be an exaggerated thought experiment where an AI doesn't actually "want" anything, it's "literally doing" what it's supposed to. It's surprising that Pinker doesn't seem to see that, or sees it and doesn't believe no version of it is possible.

Construction machinery doesn't need to be alive to kill you--just left running unattended.

The difference is that AI doesn't have a well bounded construction site.

And magical action at a distance. No thanks.

They should support Mojo which might be a great fit for edge compute.

I really appreciate seeing these dense model numbers. For a large unified memory system though I expect that MoE numbers are what people are more interested in.

These numbers could and should get much better. As an example I can run Qwen3.8-27B-MXFP4 (W4A8) on 2x AMD R9700 that gets 260+ tokens/sec to start and slows down to ~110 tokens/sec over 128k context and can do the max 256k. These are for batch size 1 and throughput goes higher with batching. This is due to speculative decoding, efficient all-reduce inter-gpu compression, and custom GEMM kernels for the specific hardware. Note each R9700 only has 644 GB/s memory bandwidth.

Basically the same as always--an extended interactive Turing test. Can they speak coherently and in-depth about the things they claim to have done on their CV/resume?

Cluould be viewed like a fancy evoultion of CHICKEN (Cheney on the MTA) that used stack for everything.

Mentally convert to downtime: 99.8 is clearly twice as bad as 99.9

Similar for LLM measures from an ideal 1.0 mark.

The whole story is a joke--Microsoft has AI?

Apparently they're cooking up MAI (Microsoft AI), and I'd seen their small Phi models listed online. Calling Anthropic a competitor is hilarious.

Yup. Could be that everyone was rushing to get all their products and demos ready leading up to it.

> i.e. whether frontier AI models aren’t generating or can’t generate new mathematical ideas. [...] I give basically zero credence to the idea that AIs are incapable of this because of some intrinsic feature of how LLMs work.

I also believe there are limitations of LLMs, but not necessarily where people think. I won't expect LLMs to be creative solvers until they can tell a novel funny joke with any recognition/consistency.

Yes. Reminds me of "Poetry is the art of giving different names to the same thing" vs "Mathematics is the art of giving the same name to different things" -- Henri Poincaré. Set of spacial facts is what's left when all the incidental embellishments are removed--pure abstraction.

It's part of the functional trio: map, filter, reduce--and half of MapReduce.

Difference being anyone can be a "script kiddie". I don't think anyone could direct an AI to proofs like this one.

I've also had this kind of topological knowingness which I couldn't name. Another comment here says "schematic conceptual" which captures what it feels like for me.

Another good explanation here is that it's all taking place but that last step of creating the visual is suppressed (as intended) while conscious as that would be hallucinating.

This is a bundle where you end up paying for parts that aren't as useful:

    - 252 GB HBM3e VRAM
    - 496 GB LPDDR5X RAM

I would rather have a system using 2x Instinct MI350P GPUs (288GB total) for much less.

Personally I'm using Qwen3.8-27B (MXFP4 quant W4A8) locally hosted on a pair of AMD GPUs (with DeepSeek Harness). It starts at 250 tokens/sec down to 120 past 128k context.

At work mostly Opus 4.8 (sometimes a GPT or Gemini 3.1 Pro). I find Opus 5 chatty/slower and Fable can venture into over-engineering itself into unnecessary complications.

Busabase[0]

> Languages: TypeScript 98.4%, Other 1.6%

Not for me.

[0] https://github.com/busabase/busabase

Seems like a random rant to me.

> What is the internet now to me? In many ways it’s something that I have to tolerate to do many things that functioned fine before. I have to use an app to pay for parking. I need to create an online account to pay for a family swim at the local leisure centre. I have to submit an online form to book a slot at the local recycling centre. I need an app to check my bank and credit card balances.

It's so much more annoying to do any of these things without/before internet. I don't do half of the other things in the rest of that paragraph. Like for phones I just get a used CAD$400 Android phone with good audio output.

I think it's largely due to psychology. If 210m is considered the best, then as they approach it they may start to tense up and choke. When the goal and possibility is known as 500m, then there's no point being concerned near 210m.

Similar here peak ~250 and down to ~120 as it gets close to 128k (which is where I set DSH compaction) though it can readily do 256k.

I just got DeepSeek Harness (DSH) set up with 2x R9700 and it's rather mind blowing that these can do actual work and quickly. Up until now I've always been evaluating and searching for better hardware/model/tweaks. This is much more than I even hoped for and considered getting extra 3090/4090. Now I can stop looking/tweaking and start using it for all the different things I've yet to discover it's good for. I do plan to also try/use Hermes and Pi. DSH is annoying that every plugin install/remove requires a restart--given that "everything's a plugin".

The way it does tensor splitting without all-reduce cost over PCIe bus wasn't something I thought was possible.

What kind of performance are you getting with 4x R9700s--what do you do with all the VRAM (batching, concurrent requests, etc)?

Thanks! Didn't expect to see this here. Exactly what I needed to run Qwen3.8-27B-Quark-AWQ-MXFP4-native.gguf as well as other experiments on one or 2x R9700's (I hope).

> Anything design-y was “too Apple” and unacceptably bourgeois.

That was definitely the case of the Unity desktop that really only worked well on netbooks. That's when I lost confidence in Ubuntu for design. Loss in Canonical on the whole came later.

The previous 'personal' AI Station I had pictured was the a16z one[0].

Each MI350P[1] in the TR Halo Station has 144GB VRAM and with 4.6 PFLOPs peak MXFP6 performance.

Four liquid cooled? Yes please.

[0] https://a16z.com/building-a16zs-personal-ai-workstation-with...

[1] https://www.amd.com/en/products/accelerators/instinct/mi350/...