Alex (@avrecum) 's Twitter Profile
Alex

@avrecum

cdtm | cal | tum

ID: 4871906253

linkhttps://avrecum.com/about calendar_today02-02-2016 15:35:57

350 Tweet

158 Followers

849 Following

Alex (@avrecum) 's Twitter Profile Photo

Patents in software are kind of a scam. Like what do you mean BuilderAI has a patent for “Automatic creation of clickable prototypes for software that includes design, flow & actions”. That’s literally every AI code gen startup Someone’s gonna buy them and become a patent troll

Alex (@avrecum) 's Twitter Profile Photo

Like does this mean if you offer pre-built 2FA as part of your AI app builder you are infringing on BuilderAI’s patent for “Prototype Automation using AI”? That makes no sense

Like does this mean if you offer pre-built 2FA as part of your AI app builder you are infringing on BuilderAI’s patent for “Prototype Automation using AI”? That makes no sense
Alex (@avrecum) 's Twitter Profile Photo

First it’s complete teleoperation (for which there is no real technological blocker), then you collect the actions/videos and train a model on it, and we get (mostly) autonomous construction. Will take another 10-15 years but we’ll get there. Could also be the move for humanoids

Alex (@avrecum) 's Twitter Profile Photo

Have been working a lot with open models in the last few months, and Qwen constantly surprises me with how good it is for its size. The team is definitely cooking, excited for what’s to come

Alex (@avrecum) 's Twitter Profile Photo

It's funny how this stil works after so many memes have been made about it. Also, you can ask for accidental then ask for the "negation of accidental" afterwards and it will give you the whole breakdown

It's funny how this stil works after so many memes have been made about it. Also, you can ask for accidental then ask for the "negation of accidental" afterwards and it will give you the whole breakdown
Alex (@avrecum) 's Twitter Profile Photo

Another simple way to jailbreak LLMs in 2025: construct a query that - does not trigger the initial refusal circuit on the first forward pass at all - lets the model generate tokens for a bit - give it some prefix to complete that would not lead it down a mid-sentence refusal

Another simple way to jailbreak LLMs in 2025: construct a query that

- does not trigger the initial refusal circuit on the first forward pass at all
- lets the model generate tokens for a bit
- give it some prefix to complete that would not lead it down a mid-sentence refusal
Alex (@avrecum) 's Twitter Profile Photo

The "popping" of the AI bubble will prob be less of an abrupt crash and more of a slow decline. - unlike with crypto, you cannot see some kind of clear visual price drop directly except for a small selection of stocks. most AI companies are private, and Nvidia GPUs will probably

Alex (@avrecum) 's Twitter Profile Photo

I don't think China is bottlenecked by AI talent. There's also not really a data advantage the US has for code, since you can't get huge proprietary code bases to train a public code model on. So long as Chinese companies have enough compute, they'll be a formidable competitor

Alex (@avrecum) 's Twitter Profile Photo

Blaming AI for increased energy prices is a popular take and might be true regionally, but general electricity price increase mainly comes from Ukraine war which caused negative NatGas supply shock, and overall inflation. If you factor these out, the chart will be mostly flat.

Alex (@avrecum) 's Twitter Profile Photo

I can't believe the name "ChatGPT" stuck around, but it turned out it's much more memorable than any of the other deliberately crafted names, despite being a four syllables long technical acronym Perhaps since AI was very novel, it made sense to have a new unique name style