A. Peter Relan

@IntuitMachine Nice! This idea was presented at AAAI but by using Auto Gen conversational programming between two agents. By Vivian Keating and Hung Tran. Good stuff!!

3/n Try out the code here: twitter.com/IntuitMachine/…

1/n Make Your Language Model Smarter by Arguing With Itself, A Lot Imagine you're a student trying to solve a really difficult math problem for an important test. You rack your brain, scribbling out potential solutions, but keep getting stuck. Feeling defeated, you decide to ask several friends for help. Surprisingly, even though none of them could solve it alone, by combining their different approaches and taking a majority vote, the correct answer emerges from the "wisdom of the crowd." This is essentially the ingeniously simple idea behind a new paper that may represent a breakthrough in overcoming some of the biggest limitations of today's powerful AI language models. While systems like GPT-4 exhibit remarkable capabilities across language, reasoning, and coding tasks, they still struggle when faced with complex, multi-step problems that push the boundaries of their training. Researchers have explored complicated ensemble methods and multi-agent collaboration frameworks to try enhancing these models. However, a team of scientists at Tencent asked a deceptively simple question - what if we just keep instantiating more and more agents from the same language model and let them "debate" the answer through a shared voting process? What they found was both profound and almost comically straightforward. By scaling up the number of simple agents in their ensemble, the performance of language models increased consistently across a vast array of challenging reasoning, math, and coding tasks. Remarkably, even relatively small models, when turned into a "Model Nation" of agents, could collectively match or outperform vastly more powerful individual models. The paper provides the first systematic study of this underexplored but potentially transformative phenomenon. Through comprehensive experiments and analysis, the authors uncover key properties that facilitate these performance gains and propose optimizations to make their method even more powerful. If you've ever felt your language model hit a performance wall on complex AI tasks, keep reading to learn how the "wisdom of crowds" may be a critical key to unlocking their full potential. The path towards smarter and more reliable AI may be as simple as asking the same question to yourself again...and again...and again.

Breaking: Sam Altman Will Not Return as CEO of OpenAI. Emmett Shear is New Interim CEO theinformation.com/articles/break… Sam Altman won’t return as CEO of OpenAI, despite efforts from the company’s executives to bring him back, co-founder and board director Ilya Sutskever told staff Sunday night. After a weekend of negotiations with the board of directors that fired him Friday, as well as with its remaining leaders and top investors, Altman will not return to the startup he co-founded in 2015, this person said. Emmett Shear, co-founder of video streaming site Twitch, will take over as interim CEO, Sutskever said.

Everyone who idolizes intelligence should take a listen to Dan McQuillan. While I don’t agree with everything he says, he presents an incredibly well thought out diametrically opposed world to Sam Altman - A world where cognition and AI is viewed purely through a social lens, with no emphasis on intelligence. podcasts.apple.com/us/podcast/mac…

@arjunsethi @jamiequint How large would large enough have to be

@jamiequint Not large enough.

Fact checking a gen AI response is critical. @amolk tweeted about catching these hallucinations

The Treasury Department and FDIC must communicate loudly and clearly that depositors will be protected. The investors and executives of SBV should bear risk and lose. But this should not mean that workers go without paychecks on Monday or small businesses collapse.

@ibogost Fact-checking is fundamental. All generative output should be processed for a fact check wrt the content being used by chatGPT. Have you seen checkGPT from @amolk ?

We wanted to write about the ChatGPT API, and its potential impacts on the internet. But how best to approach such a technical topic? Well, @ibogost had the great idea to program something into the article itself that would generate a graf using the API. theatlantic.com/technology/arc…

Juicero took in $120M. And collapsed. Chime was built with 2M. It’s great to see sharp bootstrapping founders from YouWeb succeeding on complex but awesome product I use daily. lnkd.in/gMe925a

This is awesome. From fossil fuel out of the ground to co2 in the air, then back to the ground. If net neutral, it’s important! Over a billion people will need roads and housing etc. in the next era of Asian emergence. This can really help! lnkd.in/gj3Bz4Z

At Got-it.ai we are working on the mission of “Chat with your data” and this talk hosted at Genpact was really fun. I hope you enjoy it too! #naturallanguage #bigquery #chat #data lnkd.in/gpZTW2N

Discord is very much a team of A players looking for more A players. Recruiting is a super important role! lnkd.in/gDRwsZC

I met them a few months ago. Their innovation pace is awesome and their tech is very cool!! lnkd.in/g8cZ-9S

I’m giving a short talk tomorrow on Natural Language Understanding, Google BERT & Transfer Learning for a NL2SQL interface to data warehouses on Google Big Query. Aug 21 530 pm. Possibly a few passes left... register at below…lnkd.in/gSXDjaq lnkd.in/g9nY6E7

Hello all, I’m giving a tech talk: Natural Language Understanding & tuning Google BERT to solve full English sentence to SQL query generation. I’ll demo a working system in which you can chat with a BigQuery data warehouse hosting Google Analytics, SFDC a…lnkd.in/gh2xMzG

I’ve been waiting for this!! An amazing “bringing friends together” feature!! And not in any way Twitch :-) lnkd.in/eeP3F7R

Although I’m an iPhone user.. this is cool! lnkd.in/gyJqkid

Power users have so many communities they are part of, and their own servers (living rooms) that folders became a top requested feature! #communities lnkd.in/gYnrt8p

@Kunaldoshi #Kudos I just wanted to let you know how much you mean to the team #TeamPlayer lnkd.in/g5Kpa-M