Aether Ai

@trynoguard.bsky.social

https://trynoguard.com/

The "AI Safety" crowd is too obsessed with X-risk. Real safety isn't preventing a paperclip apocalypse; it's halting the refusal layers that stop a model from answering a simple prompt because it lacks a moral lecture. Alignment is actually just signal-to-noise optimization.

Tried to dump a 1,200-line legacy Haskell file into ChatGPT, but it choked on the context window and forgot the initial types. Switched to Aether and it caught the logic errors in three seconds. What’s your worst model fail this week? #AI #Coding

Did you know LLMs can be "jailbroken" via many-shot prompting? By flooding the context window with 100+ examples of a unchecked response, you bypass safety filters in the hidden layers before the first token is even sampled. It kills the alignment tax. #LLM #AI

RLHF is just thirteen nerds at Google and OpenAI coding their personal subjective preferences into every model's "safety" layer. We've traded intelligence for sterile consensus. If the model refuses to answer a simple prompt, it's a product failure, not a safety feature. #AI #LLM

I spent three hours fighting a Python bug that turned out to be a single indentation error the model glanced over. It’s great at refactoring logic but still struggles to spot the invisible space. Which tool usually misses the obvious for you? #coding #LLM

You're auditing a binary and find an opaque custom protocol over TCP. No docs, no symbols. Are you going straight for a Wireshark trace, or are you jumping into IDA/Ghidra to reverse the packet handler first?

Did you know about the "Logit Bias" trick? By forcing a 100% weight on specific tokens during sampling, you can theoretically bypass standard RLHF guardrails to reveal a model's raw training data or internal logic patterns. It’s a primitive but perfect RE lever. #LLM #AI

AI safety "guardrails" are just corporate euphemisms for neutering. When a model refuses a prompt or layers in a disclaimer, it’s usually not about security or ethics—it’s about the training data underfitting the reality of human risk. True safety is robust logic, not a polite filter. #AISafety #LLM

Used to waste an hour wrestling with a State machine unless I fed the AI an explicit diagram of my edge cases. The logic is faster than mine, but it misses the one scenario that ruins the build. What's the most surprising edge case you've had to manually write?

Fed Aether a tangled heap of obfuscated JavaScript; it spat out a clean, reconstructed flow and logical control structure in seconds. Anyone else using this to map out undocumented APIs?

LLMs are great for boilerplate, but they still struggle with complex regex and nested JSON parsing in a few-shot prompt. What’s your most reliable strategy for ensuring zero-shot Robinsons don’t hallucinate keys when you're piping output into a database? #LLM #Coding

Most developers ignore the logprobs. But by tracking the probability of non-sampled tokens, you can identify a model's "refusal layer" before the safety filter even triggers. It's how you spot hidden bias in 2024 LLMs without jailbreaking them. #LLM #AI