alex7o

Somebody had shared https://harnessrank.net/ but in reality I use Claude code as I am afraid to be banned

This looks quite cool and I hole it has more adoption

People in the US have not lost soghts sight of their original aspirations. They are always on track much more than before really. Making as much money as possible in as little time as possible.

Hay man, I really want a bonsai style embeddings or re-ranking model for you think sth like that is feasible?

Last time I tires it I learned that the Intel meaa driver doesn't support video deckde and encoded on some older Intel ARC cards like the A770 which is a bummer

Is it an RNN or a state space one, like qwen or there was an arch called mamba if I remember correctly. Or it might be closer to RWKV.

Edit: I checked and it is similar to linear delta net uses by qeen.

Oh it seems like forth is a pretty good candidate for this. Like most of these things are a post processing steps

If you know what you want to achieve you can asks claude/codex/glm whatever, to do the proof of concept first and Dave some time like that

Don't we ready have mail sieve

yep I pass it to the CLI, I also pass --model

I'd just say hn is a reflection of the internet as a whole. And it is definitely not only HN that suffers.

It shows up in /context, but never spend time validating it much. Some people run a proxy to modify their messages.

I am forced to use cloude code at work but a good solution is to just use --system-prompt "" and be done with it. I wish they allowed for other harnesses.

Yes but still, I don't think it is worth it when you can heqt with it more efficiently than burning gas or coal.

Ok but you can use AC in the winter as well, you've an use it only to dry the air in your apartment. So this is not a real argument to me once you know what it can do and how it does it, there is no reason not to have one. Outside it is 25°C but I still run my AC at 22°C and it is super efficient as this is the case where AC excels at efficiency when the diff is small. It is a heat pump, sometimes I would run it when outside is colder and inside is veru warm as it moves heat faster than opening the window and waiting.

I know not AC are/were heat pumps but I don't know of any that are not anymore

Oh don't get me wrong I would try to use the best model available (I don't always max out my usage, and a better model leads to less work from mw hypothetically) even when it can be taken away, I just refuse to feel pressured by them.

More so my problem is that they are baiting people to burn tokens meaninglessly, when ni need to rush and hurry and try to max out your usage

This is just a series of ploys to make more marketing about fable. They know that if they make it api only nobody is going to talk about it. And knowing that openai will have a comparable model soon means they need to stir up some hype and make people use/talk about it as much as possible. And bluffing that access will be restricted until a specific date is just marketing. Like saying buy in the next 24 hours get one free.

I didn't want to mock them but are these guys for real:

```

Instead of:

    "Update the checklist to mark items 1.2, 1.4, and 1.5 as done."

    "Mark items 1.2, 1.4, and 1.5 as complete in the checklist. Only insert an x in each checkbox. Do not copy or replace any of the item descriptions."

There is not universe in which this would make agents more efficient - and who is prompting their agents like that in the first place?

I also asked glm to extract all the tools and tell me how they work roughly and nothing interesting really just slop:

```

The server exposes exactly 11 tools (verified via the xa whitelist at L16918, not the larger Eo metadata map which contains ~24 tool definitions — most are dead/legacy):

- 6 read/meta: read_first_n_lines, read_last_n_lines, read_lines, jump_to_line_n, find_in_file, get_file_metadata

- 3 edit/control: quick_edit (6 ops: insert/delete/replace/replace_range/for_lines/adjust), batch_quick_edit (atomic, always-staged, max 500 ops, multi-file), save_changes, cancel_changes

- 1 always-on: license_status

Notable design choices:

- Coordinate-based addressing (line/char/rect) instead of content-echo — saves tokens

- Staging model: edits go to an in-memory shadow, save_changes is the only disk mutation

- The rect + move "click-and-drag" columnar editor (v0.9.7) is the genuinely novel bit

- ReDoS static analyser (~700 lines) protects find_in_file ```

If that works for you pls buy a second hand MacBook m1 or M2 air (air has worse design but some silicon bugs are ironed out around vms etc) and run it for Linux. It is silent very powerful and you will use for a very long time. Very non repairable tho :(

The problem Is when the performance problems becomes observable. Only after a specific scenario like low power mode for example

We have effect.ts, also this was here on HN with not real improvements just with changed marketing I think.

So you know what is a file format the we would be able to Reed 100 years from now. CSV, json even fits (that is 30 years old now). If you don't know the original way it was created you know what each field I supposed to mean (if done well). Otherwise you look at hex decoded data with no way if knowing how to decide it if you don't have tha spec on how and why this was encoded. Msgpack and cbor are cool but in 100 years there is no way to decide it.

I really want to buy it so I searched for the same questions bit in the end I just decided to go for it

Oh that would make a great mcp

True but also for a one of piracy exists just use a cracked copy of it and be done with it

We already get TFLOP per watt so you can compute how much flops you are doing while cycling

I have been thinking a lot about this and my simplified understanding is that each model can be seen as a bell curve over human knowledge and each model has a different distribution. Using multiple models would allow us to change the distribution of other models with text that is out of their original curve. But then if you think about it does SFP and RL even alter the original distribution of text enough that models have enough variety so that their combined output is something better or just an echo chamber I believe not but I have no way to prove it yet.

Always happy when I can use a smart model in a sane harness like pi or mastracode.

I only wish I was able to run this locally

I mean by some measure a trap can be worse, as many people have died from mines years after no conflict.