vit

think & build

[vit @rudnkh]

When you decide to Europemaxx in the states.

[AIDB @ai_database]

「"1から100の間でランダムな数字を言って"と何度も聞くだけで、どのLLMか特定できる」という論文が出ています。
モデルごとに答えの偏りがまったく違うため、たった1トークンでそれが指紋になるとのこと。
たとえばClaude Sonnet 5は47に集中しがちで、Qwen3-Maxは42に集中するそうです。

この方法で、研究者はOpenRouter経由で165モデルを調査。総コストはわずか35ドルで、監査1回あたり100クエリ程度あれば「モデル名どおりの本物が動いているか、それとも裏で安い別モデルにすり替えられているか」を約9割の精度で見破れたとのこと。

さらに副産物として、ある企業が「自社開発の最上位モデル」として販売しているAPIが、実は無料公開されているQwenモデルと出力の癖がそっくりで区別がつかない、という発見も報告されています。

モデルの系譜や流用が話題になる中、こうした技術は重要になっていくかもしれません。

[vit @rudnkh]

@joaxcar I know many cases where website administrators intentionally avoided updating because the new CMS versions were not backward compatible with legacy plugins that were critical to the site’s functionality.

[vit @rudnkh]

@Kekius_Sage How do I enter iddqd and idkfa?

[vit @rudnkh]

I remember when it took weeks, sometimes months, for exploits to appear after a vulnerability was disclosed. A patch would land, the diff would sit in public view, and the world still had a grace period. People argued about responsible disclosure timelines measured in days.

Within hours of the advisory, researchers were feeding the issue to frontier models.

Frontier AI models are making it easier to turn disclosed vulnerabilities into working exploits. Finding the bug still takes talent. Turning a public patch into a working PoC used to take the same kind of talent: reverse engineering, framing the chain, polishing a payload. The zero-day is turning into a zero-hour: the timeline from disclosure to exploitation is compressing from weeks into hours.

Patch windows already got much shorter!

#wp2shell

[Z A D D Y @Zaddyzaddy]

Not anymore, we already found the full PoC @BugBunny_ai

[vit @rudnkh]

@taty902204 for what?

[vit @rudnkh]

OpenClaw replaced its setup wizard with an AI agent.

When I installed the latest OpenClaw, I thought setup was already done because it drops you into what looks like the normal agent UI. It even auto-selected Codex as the backend without asking.

More importantly, why is setup AI-driven at all? Configuration should be deterministic - pick option 1 or 2 and move on. I don't want to describe my config in natural language and hope the AI gets it right.

Setup should be deterministic, not a conversation. It's a critical step, and I want to be in control of every decision.
@steipete

[Brad Mills 🔑⚡️ @bradmillscan]

hey @grok what the heck is Crestodian in OpenClaw?

It randomly showed up today in terminal when I was trying to diagnose errors after upgrading to 5.18.

Somehow I am also back to using Claude Opus 4.7 with my openclaw after the update?

[vit @rudnkh]

They just blocked the entire domain because of one sanctioned user. Geniuses.

[domain .ME @domainME]

Statement Regarding the Suspension of https://t.co/ku1Zi3GLMi

The .ME Registry works closely with law enforcement to monitor and mitigate issues across the .ME domain in accordance with applicable laws, including sanctions requirements.

On 13 July, First VPN Service (1VPNS) was included as a sanctioned entity by the U.S. Department of the Treasury. A Telegram channel using the https://t.co/ku1Zi3GLMi domain was among 1VPNS identified infrastructure. Accordingly, the https://t.co/ku1Zi3GLMi domain was suspended.

On 14 July, Telegram provided confirmation that it had removed its links and affiliations with 1VPNS. Once the confirmation was reviewed and verified, the suspension was removed from the https://t.co/ku1Zi3GLMi domain.

We appreciate Telegram’s prompt cooperation in resolving this matter.

[Ramin Nasibov @RaminNasibov]

are you this old?

[vit @rudnkh]

I've got a new device: a Google Fitbit. Over the next few days, I'm going to look into how to retrieve raw data from the Google Health API.

[vit @rudnkh]

Orca doesn’t handle remote sessions very well. In practice, you can’t easily (at least not without hacks or workarounds) run an agent on a VPS and connect to it through Orca.

Almost all of my agents run on remote VPSs. For example, if I want to access them from Orca Mobile, I have to run an Orca Server on the VPS, which requires a graphical environment.

herdr + herdr-mirror plugin solve this perfectly.

[vit @rudnkh]

@Mr_Fireside Literally the plot of The X-Files. 😀

[vit @rudnkh]

Sometimes I use AI for translations or to help expand my own ideas — or express them more clearly. In theory, a comment like that could also be classified as AI-generated due to certain stylistic markers, such as the use of an em dash. However, I'm confident that this kind of automatic blocking is a poor approach because it carries a high risk of false positives.

[vit @rudnkh]

😏

[Polymarket @Polymarket]

BREAKING: Anthropic accuses Alibaba of using nearly 25,000 fraudulent accounts to extract Claude AI model capabilities.

[vit @rudnkh]

Check that your website has an HTTPS DNS record.

It allows browsers to skip an extra round trip and connect over HTTP/3 (QUIC) right away. Otherwise, the browser first establishes a TCP connection, discovers the Alt-Svc header, and only then switches to UDP/HTTP/3.

https://t.co/rysKbq3tj2

[Pavel Durov @durov]

Banning social media for teenagers only puts them in greater danger.

Teens are forced to switch to VPNs — and unlock far worse illegal content.

We’ve seen this before. When the Russian government banned Telegram, 95% of Russian teenagers kept using it. They just moved to VPNs.

[vit @rudnkh]

Ah, the oldest trick in politics: wrap censorship in concern for children and hope nobody notices.

Funny how these "child protection" laws always end up requiring adults to prove who they are online, while teenagers learn how to use a VPN in about five minutes.

Kids bypass the restrictions, adults lose privacy, and governments gain more control.

[vit @rudnkh]

@IntCyberDigest @DeviousPL Russia and China are dealing with this quite effectively.

[vit @rudnkh]

@pieter_haeck @vonderleyen @HennaVirkkunen Europeans after hearing they can’t use Anthropic’s AI models anymore.