On the Edge #28
Since we started in Berlin, around Jan 15th, we partnered with Quantbase 🔗 , doubled our AUM, talked and reached around 80 funds, VCs, LPs, and tens of startups and partners. I was like Bugs Bunny , first base Bugs Bunny, second base Bugs bunny..! Sales, fundraising, regulations, product, research!
Now a new phase is starting at the end of our runway; we’re waiting for VCs and incubators to reply. I'll be looking to find a job as quickly as possible (or impossible!).
The funny thing is that my friends here are looking for a job, no success yet, so I thought we were successful in breaking the European economy with our startup!
Let's see if the Berlin journey is only an episode or a whole season. This will delay our launch on Product Hunt, which I longed for, but startups are like life; you never know what you're gonna get.
We talked about AI alignment and introduced the AI alignment forum🔗 in OtE #11🔗 a year ago. Now that the impact of AI is more evident than ever, besides alignment initiatives, Partnership on AI (PAI) 🔗 structured a framework to take generative AI responsibly, not by any government but the industry leaders (Adobe, TikTok, OpenAI, etc.). Midjourney is banning certain words on their platform 🔗 . Where nearly all money in VC deals is pouring into generative AI, Elon Musk announced that because LLMs lie too much, they are dangerous. He'll start an initiative in another direction. Also, OpenAI announced that they're not planning to stay in generative AI and introduced a new plan toward AGI 🔗.
So what does all this means?
Aren't current LLMs good? Of course, they are. They're highly effective in copywriting, coding, data analysis, etc.
Is the hallucination/lying bad? yes
Can the hallucination be controlled? yes, to a practical extent.
Then what's the problem if hallucination is controllable?
1) the solution to control the model is dumb. We limit the model like a caveman. 2) there's a niche market here: to build a more general LLM that is not AGI but better than current LLMs. This is Musk's business playbook which I like. It's like he's making the electric car when others are making combustion engines. Now that the technology's use case is proven, prominent, and disruptive, you can build the more efficient one. There’s also a great post about the Waluigi Effect🔗 :
After you train an LLM to satisfy a desirable property P, then it's easier to elicit the chatbot into satisfying the exact opposite of property P.
We also discussed how Yun Lecun avoided the generative models, but Musk is not going toward Lecun and general intelligence but still a better generative model. In OtE #21🔗 we mentioned that the key component of LLMs is their training process, the famous RLHF (Reinforcement Learning from Human Feedback). The control mechanism is similar to training, so Musk is looking for an engine that is less RLHF and closer to reasoning so less text prediction. In case you're interested about my personal idea of a functional design: a system that models the learning process of exceptionally intelligent people. Exceptional talens turn their knowledge into a game environment. This small notion, from my point of view, can be a key to solving a meaningful scope of intelligence (I have long and unpublished post👀 )
Google USM, smaller than Whisper more precise and robust over 100 languages🔗
In case you haven't noticed, NFTs had a comeback recently (no relation to our NFT drop!). Bored Ape launched a suceessful game where rewards were NFTs 🔗 On the other hand I think limitations like this 🔗 which prevent AI-generated art from using regular copyright are another use case for NFTs. NFTs can be like decentralized copyright (with royalty).
Reconstructing visual images from functional Magnetic Resonance Imaging (fMRI) signals using the latent diffusion model provided by Stable Diffusion 🔗
How quant finance can help to have a better world? Small initiatives like this🔗 in universities are the best place to find out the relation. We got invited and I told them about how Eveince might not exist by September, and they replied: “Perfect match for uncertainty!“
This post used market efficiency hypothesis (MEH) to show that financial markets don’t signal for having AGI for atleast next 30 years 🔗 It was interesting to me for two reason 1) I learned from their work and their methodology. 2) In this podcast🔗 Cliff Asness was also saying that MEH can be questioned in short-term horizones. This actually means we can’t deny the likelyhood of AGI in less than 30 years because of the market inefficiencies.
On Iran's economy, in the previous post, I mentioned the government is able to control the USD exchange rate; there’s also another angel, Argentina has 15 exchange rates for the Dollar, so ..🐒 🔗
Dance 🔗
We said we want to work with fashion sooner! Honestly, in the past if you were working in robotics, you might never ended up in a fashion show. This progress over years is the result of openness. This is my hope for humanity.
Where are you on April 13th? 🔗
Neighbor practicing flute or clarinet or oboe!
اگر فرهاد را حاصل نشد پیوند با شیرین … نه آخر جان شیرینش برآمد در تمنایی
من آن خاک وفادارم که از من بوی مهر آید … و گر بادم برد چون شعر هر جزوی به اقصایی
Playlists