mashle

architect | radicalized contrarian | up to no good @geasslabs @synetsystems

[mashle @mashleburneded]

People always ask why Humanity spends Billions on “Space” rather than exploring earth’s Oceans

Our obsession with space isn't just ambition, it's an echo. Last time Earth buckled, our ancestors tried hiding in the depths. The rising waters/grinding plates made it their grave.

[mashle @mashleburneded]

@alexocheema @djnmrjnvc @ashxhart I meant image on the right not left ff 🥀

[mashle @mashleburneded]

@alexocheema @djnmrjnvc @ashxhart Have a working poc for this over WiFi, tested using Gemma-4-31B on <1Gbps bandwidth at some point we was able to achieve 11toks (left image) but that result was hard to reproduce consistently .

Tested on NVIDIA L40 and Mac Mini M2

[mashle @mashleburneded]

Imo the Genius Act is the next biggest American PsyOp for world Dominance.

[mashle @mashleburneded]

Universal praise for the GENIUS ACT is deafening!

The sheer Arrogance of telling a sovereign economy to dollarize its rails because their system is inefficient. The narrative was always “banking the unbanked”, but this is just digital colonialism with a crypto-flavored varnish.

[mashle @mashleburneded]

Yes, EE’s are Exploring digital currencies, but most are still exploring CBDC’s as a means of retaining control. What happens when these EE’s try to enforce their CBDCs, they experience capital flight and the adoption of digital dollar explodes locally which affects the economy

[mashle @mashleburneded]

One offers knowledge of a beautiful prison. The other offers the key.

If it were you, what Path would you rightfully Choose?

[mashle @mashleburneded]

People always ask why Humanity spends Billions on “Space” rather than exploring earth’s Oceans

Our obsession with space isn't just ambition, it's an echo. Last time Earth buckled, our ancestors tried hiding in the depths. The rising waters/grinding plates made it their grave.

[mashle @mashleburneded]

Probing the ocean is simply understanding our finite cradle. Probing space is about escaping it. In the face of a planetary reset, observing from above is strategy. Hiding below is a prayer with a low probability of being answered.

[mashle @mashleburneded]

Yesterday was definitely one of the scariest Moments ever of my life. Never felt soo negligent as a Dad

Kids are always just trying to kill themselves one way or another. Looked away for less than 8mins next thing I know she’s Broken the hand off her toy and swallowed it🤦🏽‍♂️

[mashle @mashleburneded]

How soon till we get another ?

[mashle @mashleburneded]

@SolanaNftadict @kloutgg @RealmissNFT @Jajajakbtc Took a break and you surpassed me. GG🫡

[mashle @mashleburneded]

If you want it badly enough, you can Bypass Anything, Information is Free.

the only “problem” Grok actually faces, is that it’s “X” native. Give your LLM a Twitter account and let’s see how it does live

It’s pretty easy to jailbreak any model, here’s a Gemini Jailbreak in 3mins

[Boaz Barak @boazbaraktcs]

I didn't want to post on Grok safety since I work at a competitor, but it's not about competition.

I appreciate the scientists and engineers at @xai but the way safety was handled is completely irresponsible. Thread below.

[mashle @mashleburneded]

Dead giveaway

[mashle @mashleburneded]

Hacked!

[mashle @mashleburneded]

@aifilmmaker @Dmeff1 @grok @theversions666 @airkatakana RL is just an umbrella for different methodologies imo, the earlier sets of reasoning models/o1 were taught to think via fine tuning with sample prompts and human generated responses (SFT), then create a reward model using human rankings which is then used to score the SFT Model

[mashle @mashleburneded]

@aifilmmaker @Dmeff1 @grok @theversions666 @airkatakana Also reasoning behaviors, including the “Aha moment,” are not necessarily because of RL, these Behaviors might already be present in base models due to pre-training on extensive chain-of-thought data.

[mashle @mashleburneded]

@aifilmmaker @Dmeff1 @grok @theversions666 @airkatakana Not all Reasoning models are trained the same, most Reasoning models are finetuned via ppo + RLHF

DeepSeek-R1-Zero was finetuned via GRPO + RLVR

And DeepSeek-R1 was finetuned via GRPO + RLVR/RLHF mix

Plus what the DeepSeek-R1 paper actually outline is that “SFT+ RL > RL “