PhilippGille

1,233 karmajoined قبل 9 سنوات

https://github.com/philippgille

meet.hn/city/de-Leipzig

comments

That was the case for early models (Llama etc), but they got much better since then. Not perfect, but good enough.

This is from Ministral 3 14B, a 2025 model without reasoning, that you can run on your PC:

> Write a Haiku involving HackerNews, and the capability of large language models like you to reply in an exact number of words or syllables.

    Silicon whispers,
    exact words in code’s embrace—
    Haiku blooms anew.

Across multiple tries it got it wrong a couple times (by ~2 syllables). But syllables are extra tricky (because of how LLMs use tokens) and the point is that for things like "summarize in 5 bullet points" you will mostly get 5 bullet points, maybe 6, but not 10 or 20, and no need for a tool that count bullet points.

قبل 9 أيامdiscuss

Currently this is for payments with stablecoins.

For Bitcoin / Lightning these kind of pay-per-request API paywalls have existed for many years already (e.g. my own from 8 years ago [1], but others as well).

Flattr [2] existed for non-crypto micropayments.

None became mainstream. I think the friction is always the extra setup on the client side. In all 3 cases the user (API consumer) has to set up a special wallet (browser extension or something for the agent) and deposit some money/crypto on the client side first. This part needs to become simpler.

[1] https://github.com/philippgille/ln-paywall

[2] https://en.wikipedia.org/wiki/Flattr

قبل 16 يومًاdiscuss

> Kimi and GLM models have coined a new term: Thinkslop. > [...] > So for now I'm happy with just two models: GPT and DeepSeek.

1. DeepSeek V3.2, V4 Flash, V4 Pro, at high or max thinking, ... when recommending a model it should always be a precise model, not just an AI lab

2. DeepSeek V4 Flash at max thinking is the most verbose model (among top models) in the AA benchmarks. See the "Intelligence Index Token Use" chart: [1]

[1]: https://artificialanalysis.ai/models?models=gpt-5-5-high%2Cg...

الشهر الماضيdiscuss

The interesting bits on how they achieved it:

> On the model side, we applied FP4 quantization

> introduced DFlash, an efficient speculative decoding method based on block-level masked parallel prediction

> On the system side, TileRT perfectly adapts to the dynamic characteristics of these algorithms

> 1000+ tokens/s output [...] using just a single standard 8-GPU commodity node

الشهر الماضيdiscuss

The blog post has more info: https://www.minimax.io/blog/minimax-m3

الشهر الماضيdiscuss

Do you mean MiMo V2 Flash? V2.5 doesn't have a Flash version.

قبل شهرينdiscuss

It's in the article:

> HTTP also allows the DuckDB-Wasm distribution to speak Quack natively! So DuckDB running in a browser can e.g., directly connect to a DuckDB instance running in an EC2 server using Quack.

قبل شهرينdiscuss

Both the original Markdown spec [1] as well as CommonMark [2] clearly specify support for inline HTML. With that you can kind of get the best of both words depending on your use case.

For the most parts you just write the regular Markdown headers and paragraphs, embed images, insert tables etc without the need for any HTML tags, making it readable in source form. And if you want to embed an SVG file for example, which the author of the article mentions as one use case, you just embed the SVG directly, and people can render the Markdown in their favorite viewer.

Let's say you're viewing a raw Markdown file in VS Code. You come onto an HTML tag, so you hit Cmd+Shift+V to open the preview and that's it.

Of course for full-fledged web pages with interactive buttons and fully customized styling and all of that, which the author shows in some examples, this is not feasible. But you can get very far when you have mostly text/images/tables and just want to add some extras here and there.

[1] https://daringfireball.net/projects/markdown/syntax#html

[2] https://spec.commonmark.org/0.31.2/#html-blocks

قبل شهرينdiscuss

On max it uses more than twice as many tokens as on high when running the ArtificialAnalysis benchmark suite, and then it's indeed the model with the highest token usage (among the current top tier models). See the "Intelligence vs. Token Use" chart here:

https://artificialanalysis.ai/models?models=gpt-5-5%2Cgpt-5-...

قبل شهرينdiscuss

Benchmarks only paint part of the picture, but it's still a decent place to start looking into recent models:

https://huggingface.co/spaces/mteb/leaderboard

قبل 3 أشهرdiscuss

When you say "Gemini", which exact model do you mean? You know there are several and they vary a lot in how capable they are? Pro 3.1 Preview, 2.5 Pro (their latest non-preview pro model), Flash 3 Preview, ...

Same with GPT-5: Latest 5.5, prior 5.4, or actually the original 5 (.0)?

You can't talk about model performance without specifying the exact model.

قبل 3 أشهرdiscuss

> C# [...] only really works properly in Windows

What do you mean with this? Maybe you are thinking of the old ".NET Framework" runtime, which only runs on Windows? Nowadays there is ".NET Core" which runs on macOS and Linux as well.

قبل 3 أشهرdiscuss

He specifically mentions that he is using GitHub Copilot because of how Microsoft bills per request instead of token.

قبل 3 أشهرdiscuss

> it is possible with some software to have everything massively cached, with the cloud doing that, with the origin server in my basement, only accessible from the allowed cache arrangement

Do you mean a setup like:

    client -> cloud(HAProxy+Varnish) -WireGuard-> basement(backend)

Or something else?

قبل 3 أشهرdiscuss

If you are interested in scale models of New York, there's a 1:1 scale model in Minecraft: https://youtu.be/ZouSJWXFBPk

قبل 3 أشهرdiscuss

The article tries to sell it to people who can't run Docker locally (e.g. locked down permissions in enterprise environments, slow old laptop), but hasn't it already been possible to use remote Docker engines?

So the news is that they're offering to host those remotes now, right?

قبل 3 أشهرdiscuss

Is this not just about extra credit? So what's included in the subscription doesn't change - just extra credits are now token based instead of message based? (For Plus/Pro)

قبل 3 أشهرdiscuss

The OpenRouter usage stats indicate the opposite: https://openrouter.ai/rankings?view=month

قبل 3 أشهرdiscuss

This is their MBP 14" M5 Max review, with a "Battery life" section and their standard web browsing test: https://www.notebookcheck.net/M5-Max-with-inconsistent-perfo...

15h 10min

قبل 3 أشهرdiscuss

The article does list what Tailscale adds on top of WireGuard:

> WireGuard by itself is mostly the data plane. Tailscale adds the control plane on top: identity/SSO, peer discovery, NAT traversal coordination, ACL distribution, route distribution (including exit node default routes), MagicDNS, and fast device revocation.