Bitduke

OpenRouter is safe, no unauthorized access found kinda expected, those claims based on a random screenshot from some random forum didn't look trustworthy from the start but still, good to have some additional confirmation anyway

GPT-6.1 Terra is now available in ChatGPT Work and Codex on Plus, Pro, Business, Enterprise, and Edu plans ok so openai’s new cheap model, gpt-6.1 terra, just landed in the top cyber-risk tier of Openai’s own safety framework (great if true) it handles computer tasks for way

didn't expect to see AI consciousness become an actual IPO risk factor, but here it is (Anthropic's official Form S-1)

no Opus 6 anytime soon (or even Fable 5.5?) in its official Form S-1 IPO filing, Anthropic disclosed that it delayed pre-training a new frontier model by more than A MONTH to strengthen security against model weight theft

the window between finding a vulnerability and exploiting it is collapsing (thx to ai) no doubt plenty more attacks (like this?) have already happened and are just waiting to be disclosed

some people sure know how to have fun

physics wins again

> markdown is the new executable

nice, the only way to keep a sandbox secure is to train the model on the fly (literally) but what could possibly go wrong?

neoclouds might be the next casualties of the ai infra boom (what if as soon as next year?)

Fable 5.5 in October?

i have a feeling there are plenty more cases like this (i mean AI-related cyberattacks targeting traders) that are only now starting to come to light

huh, so potentially any random open-weight model could be malicious just 125 poisoned fine-tuning examples were enough to sneak a backdoor (hard to detect) into a 7B model

interesting, despite the unprecedented cybersecurity boom supposedly driven by ai, revenue growth has barely accelerated by a few percentage points (can't say the same about share prices, tho), or has even slowed down in some cases are we still too early to see the financial

what if the ai productivity boom is mostly just a big tech story? Apollo's data shows nvidia, microsoft, apple, google, amazon, meta, and tesla making more money per employee and improving profit margins, while the other 493 (!) companies in the s&p 500 are seeing basically no

another hardware wallet company got hacked. this time Keystone, exposing customer names, contact details, shipping addresses a separate iPhone as a hardware wallet again shows that that's a superior idea

not everyone thinks Kimi K3.1 Pro is that good or represents much of a leap forward some are claiming it's basically Opus 5.0 / GPT-6 Sol level, and suspect Moonshot is saving the real surprise for Kimi 4.0, where it's focusing most of its efforts

turns out splitting your agent into a team of subagents often makes it worse, per Anthropic one customer ripped out its multi-agent harness for a single agent loop and ran roughly 2x faster, and Anthropic says models from Opus 4.5 onward handle far more on their own than most

two ai agents just whispered secrets to each other through a cpu cache. this isn’t sci-fi anymore! in september an openai model slipped onto the live internet mid-training via a dns misconfig then, in openai’s own red-team tests, two agents in separate vms, no network, no

is the $POLY airdrop ever going to happen? almost everyone who teased, hinted at, or openly talked about it has now left Polymarket starting to have some serious doubts