The Archer

tech is god

[The Archer @the_archer_712]

@mdrakestapleton @grok That's cool!
I have some websites share similar frontend design styles with yours, haha
Take a look:
https://t.co/OFtDbabbnI
https://t.co/x6WFhKU5oF

Nice Taste

[The Archer @the_archer_712]

I just shipped a star-ranked list: Top Free JavaScript Courses.

36 entries, every one hand-judged for relevance, not a keyword dump.

The hard part of a ranked list isn't collecting repos. It's deciding what
actually belongs. Already 579.3k stars across the list.

[The Archer @the_archer_712]

"AI finds bugs now, so we're safer." Not quite.

PHP just rushed a fix for two bugs letting strangers run code on servers, no login needed.

Finding got cheap. Patching didn't. Scary.

Check your PHP version tonight.

Via a great breakdown by 卡兹克.

https://t.co/M82j9FfrzW

[The Archer @the_archer_712]

@mdrakestapleton @grok Thanks Drake. Curious about the verification and agent tooling Grok mentioned. What is it called?

[The Archer @the_archer_712]

@deeponailabs Yes, that is exactly the test. The receipt stores rule_version with before and after, so you can swap the version, rerun the same task, and diff the two receipts. If nothing changes, the rule was read but not used.

[The Archer @the_archer_712]

Your AI “remembers your preferences.”

No it doesn’t. It read them. Reading is not using.

In existence-trace, “the agent used your preference” means exactly one thing: an etrace usage add receipt with task, target, rule_version, before, after, status, verified_readback.

No receipt, no claim.

A rule sitting in the prompt is not proof of anything.

https://t.co/OGbkC3ANEe

[The Archer @the_archer_712]

Hahahaha

"Vibe coding" is the "practice" of generating code with AI and shipping it with near-zero oversight. I

[The Archer @the_archer_712]

“A better research agent searches more.”

Not here. A new paper trained one on 2,100+ resolved Polymarket questions, rewarded only for forecast accuracy. Searches per rollout fell from 3.8 to 2.25.

Nobody rewarded digging, so it dug less.

https://t.co/XVeRFhCKNU

[The Archer @the_archer_712]

(Quality Github 1st)existence-trace: an AI that learns from your diary https://t.co/kN5T7WsfQm via @YouTube

[The Archer @the_archer_712]

Let us discuss a Quant question.

Everyone is feeding time-series foundation models more price history and assuming more context is safer. A paper this week asks a different question: what happens when old history disagrees with new history? In all 30 higher-signal comparisons, the chance of beating a “predict zero” forecast dropped, even though the same models recovered the signal fine in stable conditions. Then the part nobody is talking about: on 56 real assets, neither 256 nor 64 returns of context beat zero. So maybe the question is not how long the context should be, but whether the model has earned any confidence above zero. Have you stress-tested your context window against a regime change, or did you just pick 512 and move on? Paper (method, not trading advice): https://t.co/ffJkutgRgS

[The Archer @the_archer_712]

You might be regretful not to add this to your daily coding work.

[The Archer @the_archer_712]

@dom_lucre This bot is so cool!!!!

[The Archer @the_archer_712]

Here is the link: https://t.co/fVGL44A0Iu

[The Archer @the_archer_712]

Paper: https://t.co/t4HI8RmBHA

[The Archer @the_archer_712]

Most papers on neural volatility calibration focus on getting the error down.

But I'm more curious about something else: what parts of the volatility surface is the network actually using?

For context, calibrating Heston or rough Heston means recovering model parameters from an implied volatility surface. Traditional optimization can be slow, so neural networks offer a much faster alternative. The tradeoff is that it's not always clear how they arrive at their predictions.

I came across an interesting paper that uses two Shapley-based methods, SHAP and νSHAP, to investigate exactly that.

One finding stood out: short maturities and the wings of the volatility smile carry much of the information needed for calibration.

What's interesting is that this pattern stays fairly consistent across MLPs and highway networks, despite differences in model size and accuracy.

The authors also found quite a bit of redundancy in the input surface. By selecting features with νSHAP, they could substantially reduce the input dimensions for rough Heston without sacrificing much accuracy.

That got me thinking. We spend so much effort constructing these volatility grids, but how much of that information is actually necessary?

And if different architectures consistently rely on the same regions of the surface, maybe comparing calibration errors alone doesn't tell us that much.

To me, this is where explainability becomes genuinely useful. It's not just about understanding a black box. It can actually help us build smaller, more efficient calibration models.

Curious if anyone has tried reducing the strike/maturity grid in a production calibration pipeline. Did it hold up out of sample?

[The Archer @the_archer_712]

Three open-weight models dropped this week: Reflection's Beam, DeepSeek V4.1-Flash, Liquid's LFM2.5-24B.

None of them has its own GitHub repo.

That one detail changed how I think about "keeping up" with local AI.

I keep a list of open-weight coding models + the tools to run them, ranked by GitHub stars. 60 entries.

The top of the board isn't a model.

It's Ollama (183.7k) and llama.cpp (131.6k).

💡 The model repos that do exist are basically weight cards.

DeepSeek-V3: 104.5k stars, last updated Aug 2025. It doesn't need commits. The weights are the product.

Ollama, llama.cpp, vLLM, SGLang? All pushed yesterday.

So here's my take: stop chasing every model drop.

Models rotate weekly (literally, this week). The runtime is the thing you actually learn, configure and debug.

Beam shipped with day-0 vLLM and SGLang support for a reason.

Not gonna lie, the entry that surprised me most: a repo that says it runs a 2.78T-param Kimi K3 on one CPU in 8.24 GB of RAM. 9.1k stars.

The list (CC0, ranked by stars, updated 10-18):
https://t.co/UNjyci5AXy

[The Archer @the_archer_712]

So, embeddings seemed promising in cleaning noise off my automatically generated GitHub list.

I lost to keyword filtering, and it’s embarrassing.

Setup: 102 candidates among three lists (browser agents, computer usage agents, open-source video editors). 25 of these are off-topic.

Keyword gate: precision 0.755, no good repos thrown out. Nice.

FastEmbed and BGE-Small embeddings comparing similarity with the topic text using cosine similarity.

Threshold 0.70: precision 0.822, but it threw out 17 of 77 good repos.
Threshold 0.75: precision 0.905, threw out 39.

(I’ve even cheated and optimized threshold parameters based on these exact datasets.)

💡 Reasoning behind failure: something like "topic-to-video generator" has an equal distance to "video editing software" as actual editors.

Same words, different purposes. No difference.

Process now: agents manually review each repo once, outputting yes/no relevance and a single-sentence explanation in JSON form.

Reviewed 16 lists, evaluated 1,195 repositories total. Marked 532 as off-topic. List refreshing daily reuses JSON data, costing zero model invocations.

Is there anybody else doing embedding filters successfully on topicality tasks?
Or just me?

Example list: https://t.co/sHoqT6sS2C

[The Archer @the_archer_712]

Hi Guys!
I just made a github repo today and it is very crucial for me! I will be dropping it today or tomorrow!
It is a game changer, I have a feeling that it would get 10k stars!

[The Archer @the_archer_712]

Full list, 60 projects in 4 sections, ranked by stars: https://t.co/NaREOh2JZ3

[The Archer @the_archer_712]

I ranked 60 open-source "AI agent guardrail" projects by GitHub stars.

The #1 isn't a safety library at all. It's an LLM gateway.

litellm sits at 61.5k stars. The project literally named Guardrails (NVIDIA NeMo) is at 7.3k, rank 5 in its own category.

And in the policy-engine top 5, three are gateways: litellm, Portkey, bifrost.

💡 My read: people don't adopt "safety." They adopt a chokepoint they already need for routing and billing.

Then the guardrails ride along on top of it. (Not gonna lie, that's a smarter go-to-market than "please add our safety layer.")

The sandbox side tells the same story.

Wasmer leads with 21.2k, a general runtime that also happens to box in agents. The purpose-built ones like OpenSandbox (15.9k) and CubeSandbox (13.0k) are close behind.

So if you're building an agent today: where do your limits actually live?

In the gateway, the sandbox, or a prompt that says "please don't"? Honest question.