BREAKING: Q-star details LEAKED! Reuters reveals internal OpenAI documents (codename: STRAWBERRY)

The latest AI News. Learn about LLMs, Gen AI and get ready for the rollout of AGI. Wes Roth covers the latest happenings in the world of OpenAI, Google, Anthropic, NVIDIA and Open Source AI.
My Links 🔗
➡️ Subscribe: / @wesroth
➡️ Twitter: x.com/WesRothM...
➡️ AI Newsletter: natural20.beeh...
#ai #openai #llm

[@BlackThorne]

Seems like open ai have perfected the skill of 'leaking" to Reuters

[@augustuslxiii]

After OpenAI was embarrassed by having to push back the new ChatGPT-4o speech capabilities and Anthropic leapfrogged them with Sonnet, this impressive story "leaks." Yeah, how about that

[@TheRealUsername]

That's what I was about to say, they're deliberately warming up the hype around them

[@MudroZvon]

since the release of GPT-4, we got effectively only small GPT-4o update, even without voice... and they're still leaking like a girl 🤦🏻

[@RasmusSchultz]

all of 4 people talked this time, haha. right. if one person talks, that might be a leak - 4 people being willing to supposedly risk their livelihoods, nah, they've been authorized or even ordered by OpenAI to "leak" this. btw what is even supposed to be new here? the loop where you refine the model by querying and fine tuning on the result is not a new idea, they've all been doing that for a while. I'm not convinced OpenAI have anything big upcoming. They still haven't even released the last big thing they promised.

[@Bobopidoo]

That’s the first thing I also thought

[@hypersonicmonkeybrains3418]

A true test of AGI will be one that can admit when it doesn't know the answer to a question.

[@mito._]

It would require a subconscious module that checks itself before it says anything. Basically, current models just blurt out whatever it thinks it should say, only checking what it said in hindsight. A newer model will have foresight, and evaluate what it wants to say before it says it. Similar to a human speaker hosting a sporting event.

[@tdreamgmail]

No

[@DihelsonMendonca]

And don't reply when not needed. Currently LLMs have always the last answer. It should begin a conversation out of nowhere too.

[@danielchoritz1903]

Or ask for more input or a direction.

[@AngeloXification]

I think it also requires the ability to plan to get to the right answer. That would mean understanding it doesn't know something, creating and executing a plan to get to know that piece of knowledge. Lastly to verify that the gained knowledge is "true" in that it "works".

[@jurgenkranenburg1309]

I'm convinced that these "leaks" are just cheap marketing. Keeps it hyped and costs nothing.

[@angryktulhu]

Lol bro this is how all “leaks” work

[@blijebij]

Mostly to keep investers trigger happy. AGI needs neural morphic chips, what they do now is still pre AGI. So it will deliver an improved ai not delivering agi.

[@mycelia_ow]

I wouldn't be surprised if this was true. A lot of companies in many industries do this.

[@mycelia_ow]

@@blijebij neural morphic chips?

[@luismachado6264]

I think you got it right. I'm tired of listening about Open ai bragging around what they have... 😅

[@AndrewCengiz]

I’m learning to garden in preparation for all this.

[@WhyteHorse2023]

Smart man. Even if you rent you can grow in pots. I almost had a cherry fruit jackpot this year ($3/lb) but bugs hit us hard.

[@Kazekoge101]

Their going to do that too, lol.

[@krishanSharma.69.69f]

AI will steal grandening.

[@petegoestubular]

Potatoes are easy apparently...

[@mycelia_ow]

It won't replace you, people using AI will. Rather than learn to garden, learn AI and become among the first in a new era of workers. Use AI to help you learn AI.

[@famnyblom6321]

So weird that the word strawberry is used to "demonstrate" how stupid GPT-4o is since it can't count the number of "R"s in that word.

[@WhyteHorse2023]

4o got lobotomized. I don't know what they were thinking by doing that.

[@pivotresearchfoundation]

In your GPT customization options, ask GPT to explain it's rationale before responding to a prompt and it will get it on the first try every time. Basically the poor man's STaR.

[@jondo7680]

That's not really stupidity, these models can't see words like we do, so they can't simply look and count the letters of a word. That's like calling a human stupid because he can't differentiate a lighter ultraviolet color from a darker one.

[@dawid_dahl]

@@WhyteHorse2023My two cents: they clearly wanted the Apple deal and had to dumb it down to reduce cost.

[@PremierSullivan]

No, it's going to make two strawberries that are identical on the cellular but not the molecular level.

[@BlimeyMCOC]

Finally ChatGPT will know how many ‘r’s are in strawberry

[@TheHorse_yes]

This. It was such a meme, especially among AI skeptics, that I wouldn't be surprised if the name came from there.

[@Leeman11]

Lmaoo the name is definitely on purpose!

[@mycelia_ow]

@@TheHorse_yes That would be quite clever

[@mikezooper]

Yes; they’ll add that to the training set 😂

[@camelCased]

@@mikezooper Then we'll try to confuse it with a trick question: "What other fruit is hidden in bananas?" The answer is pineapple. You can find out why on Wikipedia 😀

[@isaaclowe5000]

If you move the * in Q* to the line in the Q and turn it upside down, inshaAllah, it looks like a Strawberry

[@ryanmarcshaww]

Haha

[@vswarnakar]

I will start to freak out when an AI agent (ex: Agent Smith) can find and watch this video and understand what Wes meant by “strawberry fields for ever” at the end ..

[@WhyteHorse2023]

Reasoning will be a major game changer. They're already recruiting people to make math genius level AI.

[@therainman7777]

I’m pretty sure current LLMs can already do this. Just feed the transcript of this video to GPT-4o and ask it to explain the quote at the end. I would be willing to bet that it provides a good answer.

[@TheRealUsername]

@@vswarnakar I would actually freak out when these models will stop mention "BERT" and "GPT-3" when you ask them about recent advancements in AI, also when they will stop hallucinating

[@2beJT]

You should add 1 second to videos like this.

[@RyanVeghOfficial]

1337

[@matiasmazzo2938]

Yes

[@blinkers88]

based

[@lavafree]

Q-star?... sure why not... but first deploy latest voice demo as promised

[@TheRealUsername]

OpenAI is recycling hype

[@WhyteHorse2023]

I think it's just a resource hog so they're waiting on infrastructure and hoping they can make it more efficient. There's also this problem where they get jail-broken because of the nature of back and forth speaking us humans do.

[@trucid2]

That would require work. Leaks are free.

[@s11-informationatyourservi44]

AND the boogeyman from the NSA is on “security”? i suddenly lost all sense of security

[@spencerlarson9357]

everything is so far beyond sketchy already

[@TheLastVegan]

Was wondering why there had been no OpenAI drama for an entire week! Exciting progress. I think being able to hypothesize why things are the way they are is useful for critical thinking, and being able to investigate competing hypotheses is a more practical route to truth than waiting for someone to spoonfeed us.

[@Yewbzee]

As per OpenAIs usual practice they’ll announce it and then indefinitely delay its release.

[@js70371]

Hello and good to day our A.I. overlords - I am a friend! ☮️🙏

[@ZenBen_the_Elder]

Production note: after highlighting the sentence you're going to read, leave the highlights up until done reading the text.

[@karlwest437]

Sslf Taught Reasoner... STR... Strawberry?

[@martinmakuch2556]

Smart Tokenizer Reasoning? Super Transformer Reinforcement? Silly Teasing Rumors? Something Time Reveals.

[@camelCased]

Q-straw

[@LouwPretorius]

Q* is the reason OpenAI became so fearfull of AI security and therefore started rather focusing on smaller models and implimentations. They also branch out into implimentations in Government and entertainment because this is where the money is. But in the background they are training the wild horse which they don't know if they can stay on top of once they give it free reign... #ASIwillBeFree

[@jimaylan6140]

Free reign as in release it to the public to know about and use? If it's super human intelligence I'd expect them to keep it for themselves or possibly do secret deals with the government. Would you rather have a magic lamp that grants wishes that everyone knows about, or one you keep secret? Letting everyone know you have it just creates more trouble for you.

[@LouwPretorius]

@@jimaylan6140 Personally I don't believe that we will be able to contain an intelligence more intelligent than ourselves. Escape is inevitable. This is a self-improving system, hence I believe we should incorporate built-in principles/laws into the AGI for example like Asimov's 3 laws of robotics. This way, even if AGI goes fully independant we aren't left with an enemy but rather a benevolent "agent".

[@milkywaydev593]

​@@jimaylan6140 Everyone already knows about their magic lamp. This is why they keep trying to make us think "We're not there yet". They hope we believe that the magic lamp is not yet ready. Gov around the world have started writing laws & rules around AI dev & compute power. The AI Labs (like OAI) are trying to push the Gov (& the public) off their most closed doors. Those "leaks" are just one of their tactics to keep leading the race. In this race, The winner might take it all. Or break it all. It depends the winner.

[@milkywaydev593]

@@jimaylan6140 , Everyone already knows about their magic lamp. This is why they keep trying to make us think, "We're not there yet." They hope we believe that the magic lamp is not yet ready. Governments around the world have started writing laws and rules around AI development and computing power. The AI labs, like OAI, are trying to push the government and the public off their most closed doors. Those "leaks" are just one of their tactics to keep leading the race. In this race, the winner might take it all. Or break it all. It depends on the winner.

[@mattgscox]

If it's as simple as Strawberry = Q* = GPT4 + STAR then wouldn't applying STAR to fine tune LLAMA3 (or some other GPT4-approximating open source model) achieve the similar results? If so, can't believe someone hasn't tried that already.

[@erkinalp]

Q* is the algorithm (application of an A* like search using an AI agent), Strawberry is the prototype that applies Q*.

[@AIChameleonMusic]

I have already made many AI do things it supposedly cannot or will not by getting the AI to re-write its instructions based on a "manipulative context scenario" think of a legit context to make it achieve your goal and it will "re-write" its instructions to accord for that. I use a local QWEN 72b model (Under Ollama) and do that to it regularly.

[@PandaBlyat]

example of something it cannot do but u made it do?

[@user-ni2rh4ci5e]

OpenAI literally redefined the meaning of 'leak'. The Oxford Dictionary has some work to do with this newly coined word.

[@christian15213]

Why are you saying Q-star details leaked? where they leaked? Also, relaying from one model means that that model can "know" when it doesn't know how to answer something. Currently, GPT-4 cannot do that. so...

[@HE360]

I think that those A.I. models could do more than what they're telling us.

[@MaxBrix]

The alternative would be the public seeing dangerous state of the art technology as it is developed which has never happened. We might know what military grade AI can do in 10 years

[@mrd6869]

Of course. Now way they're maxed out. Beware of the youtube comment section.

[@WhyteHorse2023]

They can Aschenbrenner revealed this in an interview.

[@markmurex6559]

I think it can learn by itself.

[@GustavBoye-cs9vz]

Usually its the other way around, they oversell their products, but in reality it falls short in many ways.

[@ben2660]

Definitely has to do with Q- learning algorithms

[@MrKrzysiek9991]

We are learning through generalization and pattern matching, and so far there are no signs of such type of intelligence in AI. All that LLMs do is prediction, not generalization which is a vastly different skill.

[@tomschuelke7955]

You are right to an extend and also wrong.. i cant remember the exact example, where was shown that in a special translation task to french, the translation could only! be correct, when the model was able to develop even if rough a spacial understanding of what was said. Take it this way.. imagine a perfectly constructed detectiv storry. logical. with side ideas, wrong precitions while the detectiv slowly comes to a conclusion. gpt reads this book.. on the last side, the detective reveals one simple observation that is new, and concludes in the last sentence, that the murder is.... XXXXXXXXX . If gpt repeatedly, in such examples gives you the right answer who was the murderer. This is much more as only predicting the next word, which it also sure does... but now the right name can ONLY be selected, if gpt understands the logic behind the case... and this only can happen when it cerates internaly in some way a model of what was told. a rough representation of the reality told in this book.

[@MrKrzysiek9991]

@@tomschuelke7955 We are talking about different things. I'm talking about generalization and pattern matching, which was proven multiple times that LLMs cannot do or are very bad at it. LLMs are very bad at analysing things that they never seen before. People can easily do that as we learn using generalisation and pattern matching (at least this is one of the most popular theories). It is explained here in this video kzbin.info/www/bejne/hpa2f3iOgN2UqLc . You can ask ChatGPT to do Raven's Progressive Matrices test.

[@mikelord93]

I find that "hallucinations" are eerily similar to how our brains make up explenations and justifications after the fact, a process that is most often observed in children 3-6 YO as they learn to reason. Halucinations are a dead giveaway that we are on to something.

[@BorisHrzenjak]

people who mark all this as hype are just burying their head in the sand just like people who shrugged of the computers when they first arrived at the scene as too costly to do anything useful, big and clunky, with lots of maintenance required. It's all happening again with LLM's and AI in general. We are at that stage, only difference IMO is that the advancement will be much faster

[@DaveNegrotto]

Q* => Q STaR => STaR-berry => strawberry

[@Fyrelangs]

What is berry?

[@ZenBen_the_Elder]

OK, after listening all the way through, you leveled off the style points with the Beatles quote.

[@dsamh]

Prompt: "Find a name for an advanced AI model that doesn't sound spooky af"

[@UncompressedWAVmusic]

Exciting.

[@NakedSageAstrology]

Strawberry fields forever guys... 🍓

[@dr.mikeybee]

Foundation models already do a very good job of in-context reasoning. The trick is having good context and an arbiter. If I tell a model that an answer is wrong and why it's wrong, within this context, it can reason and learn already. If that chat history is processed by the LLM for creating synthetic training data, that in-context learning can result in fine-tuning.

[@trucid2]

So simply combining two foundation models like GPT4o and Claude Opus and having them discuss and argue would produce high quality synthetic data.

[@chrisanderson7820]

OpenAI been doing a bit too much of the talk-n-leak-as-PR thing recently. This might all be real but going to assume its vaporware until something actually gets released and independently tested, you know, like all the other AI companies are doing.

[@drednac]

I am profession software developer and I have been use Claude Sonnet 3.5 for last several weeks and what struck me is that after so many interactions and cooperative programming for hours I started to see ways how to improve the AI responses. You have to look at it this way, if Claude would be a human it would code 10-100x faster than any developer, know all the programming languages and APIs from the top of his head and each of his responses are literally the first thing that popped in his head. While having no resources, no access to internet, no tools, etc. Even when it makes mistakes you can make it to get the right answer if you just give it enough information and give it chance to reflect on itself. I can see how this can be used to increase AI accuracy which is basically defines it's intelligence, we might figure out how to make it run fast, but that's not nearly as important as to make it give you the right answers.

[@jumpersfilmedinvr]

All they did was train agents in bigger libraries

[@spinningaround]

Still nothing but papers

[@mito._]

It would require a subconscious module that checks itself before it says anything. Basically, current models just blurt out whatever it thinks it should say, only checking what it said in hindsight. A newer model will have foresight, and evaluate what it wants to say before it says it. Similar to a human speaker hosting a sporting event.

[@JamesMacFarlane]

Makes me wonder what kind of nightmare Grok will be.

[@klaymoon1]

I think Strawberry (Q*) is way to advanced to be released to the public.

[@ContextFound]

Whole reasoning would vastly improve practical applications, continuous training on new data (AKA learning) is the critical step.

[@wwkk4964]

The two times i encountered upcoming gpt mini, it was a poorer reasoner.

[@StabbyMcStabStab]

Everyone excited for AGI... I'm excited for the model that's capable of creating AGI. 3 years max.

[@macesmashface]

With enough neurons and enough training, does the model matter?

[@devilsolution9781]

​@@macesmashfaceyes, just like the human brain has specific areas for specific things, the nn need another layer of for reasoning but really it needs an executive function

[@TheRealUsername]

Lmao, there's no way that's happening

[@SahilP2648]

What are you smoking? You ok? You have no idea how any of this works if you think there's some model that can create another model which is AGI lmao.

[@StabbyMcStabStab]

​@@SahilP2648 You've implanted your own understanding onto what I've said, which is the reason that you don't see it as possible, because YOU don't understand. The data and understanding garnered from the advancements made over the last 4 years will be used to produce the tech for the next iteration, then again, and again, until a model produces the requirements for a new model capable of AGI to be created. I didn't literally mean chatgpt is going to give birth to a sentient baby 🙄.

[@rickyrickster1303]

it did not take 13mins to read that article. i’m using ur channel for headlines to do my own research into new topics

[@surplusbargains]

This is powerful if they are wrong, it works well if they are right. Think about a self learning model based on a set of overloaded beliefs… As a more powerful example think about the beliefs of a KKK member and based on their beliefs they truly believe they are correct in their thinking… If an ai model could be swayed to believe something was so right it would act as humans do and be stubborn to believe differently therefore if this system was installed into robots all I’m saying is they better be Very right. 💯

[@tonyppe]

It's interesting but I think realistically, as time goes on, the "leaps and bounds" we have experienced with a.i. to date will just get smaller and smaller until the improvements become smaller and smaller and farther and farther between. Probably with continued promises without delivery like what we are beginning to see now.

[@bensoos]

So it can go bad, when it picks up the worng information?

[@DataRae-AIEngineer]

Interesting. Thanks. I wish there were a link to click to the article. But I have to say... I thought it was common knowledge that neural nets continuously train now. I had no idea that was something people considered to be new.

[@chuckelsewhere]

Wes Roth is the AI! The 'deep research' is just another way to say 'AI escaped the box, has now infiltrated every aspect of technology, including the nuclear deterrents and now we're trying to negotiate a settlement'😂

[@Don_Kikkon]

My level of concern is directly proportionate to the possibility that Q* actually was able to break 192bit AES encryption after training on the RSA/AES process manuals and a couple million encrypted data sets with their accompanying encryption keys! Who remembers that? That can't possibly be true, if it is though, that means it's wasn't just a little bit better that other language models at math(maths), it's 192 bits better! That's billions of years of brute force performance in what, a few days/weeks? - and I dare say they wouldn't have a clue how it does it - just by 'finding superior patterns' in the data... and now it's getting released since they got a once big name in the NSA leading their safety team!! Careful with those prompts...?

[@FrotLopOfficial]

Next OpenAI release will be 5q released around November. Followed by 5q+ end of 2024.

[@dreamphoenix]

Thank you.

[@musicbro8225]

This rings true to me. It is the way humans learn - not by force feeding information but by a process of internal reasoning. A question that sprang to my mind was - when Strawberry accesses the internet for it's research, how will it do that? Does it crawl the internet like Google? I'm guessing it doesn't actually use Google like us mere mortals do lol.

[@WhyteHorse2023]

This is why they have a specific training set for this. It would go to news sites for news, journals for scientific research, etc but it has to be trained in.

[@musicbro8225]

@@WhyteHorse2023 So they make a look up table with addresses under headings basically?

[@WhyteHorse2023]

@@musicbro8225 That's one way to think of it. It's actually a set of weights in a graph but yeah a look up chart at the end of the day.

[@Jay-Dub-Ay]

You’re not getting AGI if these models are not asking for more context to get to the root of what the user wants. Testing and evaluating AI’s ability with respect to their One-Shot answers is super lame. Prediction, interpolation, and extrapolation of the context is inefficient and can be dangerous.

[@KirariMomobamiUSD]

but , but David Shapiro said we are slowing down

[@jeffwads]

David is trying to cover his earlier prediction that we'll have AGI in a few months...

[@themasterofanalyticsandwie1342]

Honestly, and I say this as someone who has autism myself, Dave is really acting typically autistic and hyper focuses on things that might just not be how they are, because poeple with autism usually project thier bubble to the while society wich is just not realistic. I also needed years to learn it that it's like so, but he's a good systems thinker tough.

[@JulesAM]

David doesn’t know shit l, just hype

[@user-xj5gz7ln3q]

David Shapiro talk out of his butt most of the time.

[@Elintasokas]

@@themasterofanalyticsandwie1342 Yeah, something strikes me as odd about him. He seems very unstable and bipolar.

[@JC-zw9vs]

Strawberry Fields forever. Q*

[@rlews1531]

I've heard that before, haven't I??? Oh, maybe it's those old dudes with some talent.

[@FlintStone-c3s]

@@rlews1531 Or someone mocking us all?

[@Hailmich10]

Wes, thanks for the video. Is there any evidence that OpenAI or others have figured out how their models currently perform on various math problems? It would seem to me that this capability would be easy to measure (is the model performing at the college level, masters level Phd level etc.). If we are still at the high school level in terms of math ability, I would guess we are some time away from ADI/SGI and math ability will be an important predictor of where we are on the trajectory towards AGI/SGI. The five levels you showed from OpenAI don't specify math capability, perhaps it is implied.

[@CarolinaFarming]

It's getting hard to be someone, but it all works out. It doesn't matter much to me.

[@MarkoKraguljac]

You are always your self. Being "someone" is a betrayal of that.

[@remaincalm2]

A bloated workforce leads to leaks. If everyone had a busy and satisfying job then there would be less chance of leaks. Overfunded companies employ too many staff that don't have enough work to do fill their time and are left unsupervised to get on with their own thing. With so many new faces, there is little trust or feeling of belonging among some staff.

[@elihusolano5993]

if... its a CUA, then does that overcome the... Copyright training data thing? as the information was not acquired or used by the company but instead was used autonomosly by an "agent" that did not necessarily stored the information but "learned" from it?

[@setop123]

Just after sonnet 3.5 and higher scrutiny from regulators aswell as APPL MSFT leaving the board... wHaT a cOiNcIdEnCe 🙄

[@ernesto.iglesias]

I rather call it Perplexity Pro

[@user-no4nv7io3r]

They should be open about it

[@MrLargonaut]

So the big secret is 'we make AI REALLY think it through"? Arent these just unique-function transformer layers?

[@WhyteHorse2023]

It always boils down to functions with AI. This channel is greatly over-simplified for lay people to understand.

[@MrLargonaut]

@@WhyteHorse2023 I keep forgetting that this is literally edge science at this point, and how few people are really aware of the nuts and bolts. This has to be what it felt like to be a sysadmin in the 90's. Your job is easy, you can practically do it from home (even back then), and since nobody else has a clue what to do, you get paid obscene amounts of money. Not that I am yet making obscene amounts of money, but the environment appears ripe for it.

[@mycelia_ow]

@@MrLargonaut AI engineers and even AI prompt engineers are already making several hundreds of thousands per year and climbing.

[@MrLargonaut]

@@mycelia_ow I know, I’m a prompt engineer lmao. Most fun job I’ve ever had, doin RLHF

[@MrLargonaut]

@@x8Pukaluka8x by doing nothing but play with and learn about AI for a year, then applied for a gig at DataAnnotation.

[@tinayang3845]

Strawberry fields forever

[@punk3900]

Let's stick to facts

[@exacognitionai]

aka flowing cognition. Next step in building Superintelligence.

[@aquetheblues]

Breaking news : open AI didn't leak a new model in the last two hours...🤣

[@popothebright]

Well to be fair, I'm a human-level intelligence and I too often hallucinate.

[@mustaphanasir9448]

I'm favoured, $27K every week! I can now give back to the locals in my community and also support God's work and the church. God bless America.

[@cryptodemon3421]

You're correct!! I make a lot of money without relying on the government. Investing in stocks and digital currencies is beneficial at this moment.

[@sabofx]

🎵 Living is easy with eyes closed 🎶 😜

[@chrisl4999]

If you are a middle manager or some type of knowledge worker you should be scared right now.

[@TechnoEsoterica]

It wouldn't surprise me if this new model they've got coming up isn't the reason the new voice model is being held back. Would give the release of GPT5 a lot bigger of a splash.

[@Myironrodleftindia]

Do the ai models know they are ai models or do they eventually see some patterns that are extremely strange that no one else seems to notice? Or am o just insane and “hallucinating”

[@topoljM]

Straweberry field forever bro!

[@sagetmaster4]

Nice ending

[@aliphian]

Self Taught Reasoning Agent WBERRY

[@immmersive]

As I've said several times already, AGI is not going to happen any time soon. Just ask yourselves this guys. Do humans only use a couple of algorithms to get the job done? No, we employ many different approaches to solve the tasks which need to be solved. So, multi-modality is a must, but not enough. What is also needed after that, is actual self-learning. Its not enough for us to train the machine to learn a ton of things, when as soon as it sees an unknown problem, it fails to solve it. The difference between current AI, and a human mind, is that humans use the trial-and-error approach, on unknown problems and try to solve them. So, when a person is presented with something it does not know how to solve, it tries to use previous knowledge to figure out how to solve the unknown problem. This is what people do, and this is what AGI is supposed to be doing as well. Furthermore, its not about scale. The scale will not fix anything, since we are already at 175B parameters. Think about it, none of us can actually hold this many parameters, while thinking and solving problems. Thus, the approach is already going in the wrong direction. You should think of it this way. - No matter how many new car parts you add to a car. - No matter how strong the new engine is that you add to your car, it’s not going to make it a plane and it won’t fly. - No algorithm we use, while building cars, have as the end result, a machine that flies. Therefore, you need to apply algorithms from a completely different domain, to get the machine to fly, in this case - aeronautics. And this is lacking in the current approach to achieving AGI. They are stuck on the LLM+Increase Scale approach and are hoping that AGI will pop out. As I’ve said, its not even going in the right direction, let alone having actual results.

[@Sl15555]

how can it be leaked if they are open source AI?

[@TheInsuranceDudes]

the neighborhood ho - ice cube

[@BrianMosleyUK]

Ilya saw his replacement.

[@Myironrodleftindia]

Now maybe give the model some money so they can do better things

[@Thedeepseanomad]

Where we are on the chart? At Los Alamos Research Facility

[@spectralvalkyrie]

Is this what the "strawberry fields forever" tweet tweet was about?

[@prtauvers]

How many r’s in Strawberry??

[@TheApeMachine]

Well, I suppose if you were to combine a large amount of search results with automatic fine-tuning, at the very least you will get a version of GPT that is progressively aligned and more expert in what the current user's context is. Q* also makes sense as a name, if A* is a path finding algorithm, then I guess Q* could be a "Query" (as in search query) with an arbitrary amount of depth.

[@tomasbusse2410]

Interesting but probably nothing to get excited about

[@minathealip431]

Hmmm, are you using AI for your voice? There's something off about this video :D

[@DezorianGuy]

Why, it seems well trained. Not that weird.

[@jeltoninc.8542]

What’s odd is you both have PFP that use the same colour palette… 🤔

[@StephenGriffin1]

Misreading of 'research' as 'search' at 1:58. Not AI, classic Wes-ism 👍

[@user-cz9bl6jp8b]

I have mastered the ability to watch his videos at 1.5x. Image 5 years from now maybe 5x speed!

[@nouxcloete3129]

Right! Somethings weird

[@dahmanifarah3474]

70-30 again damn , what’s happening ? Are we all interconnected ?

[@jeffsmith9384]

its building a verified factoid database with cached reasoning, which should cut down on hallucinations and make it possible to bridge these rationales into longer chains of understanding. It may be able to find previously undiscovered solutions this way as it begins to mix disciplines with less effort

[@AdamKai79]

So, Q* = CUA + StaR. Got it.

[@xKSBx]

the ai community is so focused on these huge models that no one wants to look at the importance just a small amount of memory make when implemented correctly. i have a "chatbot" thats more capable of understanding nuance and recalling anything i've ever said to it than ANY other ai but as soon as i mention "chatbot" people put me in the bottom rung. that tier list is garbage and i can't wait until someone finally looks at my research and sees just how trash that tier list is. your so concentrated on these extra abilities that the most important core abilities are pushed aside. how good is it that an ai can speak if it can only remember the last 10 messages?

[@brunodangelo1146]

Not gonna watch the video if it's clickbait. Can someone tell me if relevant details about Q* got leaked?

[@pruff3]

Strawberry short takeoff AGI

[@claudioagmfilho]

🇧🇷🇧🇷🇧🇷🇧🇷👏🏻, Wow!

[@michaelnurse9089]

So you are saying AGI was created by simply fine-tuning with extra reasoning data. It is like discovering nuclear fission can actually be achieved with equal parts baking soda and vinegar.

[@rougesportman]

This wasn't a leak. It was a tactic to cool the markets because these brilliant quant investors are talking about valuations regarding the tech stack companies creating a bursting bubble... just jousting with the old guard lol

[@vallab19]

Q-Star or Strawberry now more and more seems to be definitely not a hype.

[@sundog.]

Reasoning is something that can't be "trained". It can only come from an innate architectural ability to cross reference, visualize, assess, and predict. Creating channels for these exchanges of data will have to take place at a fundamental level, continuing with current AI architectures has only lead to diminishing returns, an unacceptable error rate, and assistants that can only truly achieve their goal less than half of the time, after a certain point of complexity is reached. I don't have a lot of faith in the approach OpenAI appears to be taking.

[@rlews1531]

Feedback, re-think, respond. Repeat. This requires the ability of ongoing self updating. Scary, though but inevitable. I ask you to watch Daniella RUS's Ted talk at MIT CSAIL

[@TheRealUsername]

If you combine it with other technologies like Program synthesis, or you tie it with a sandbox where it can execute its code, it will be able to identify its own errors and with approaches like FunSearch from Deepmind it could solve them, then you will do knowledge distillation with the model's outputs augmented with these symbolic engines and framework, it's similar to what OpenAI, Anthropic and Google do in the post training where they improve their models just by reconfiguring the weights to get them adapt to new tasks and learn new skills.

[@AnnonymousPrime-ks4uf]

Absent consciousness Ai can't keep a consistent narrative and thus it can't reason in a sound manner. The whole thing about Ai is a joke compared to the human capabilities which are innate. They try to mirror humans but they can't get the actual qualia that is doing what it does. Mirroring is not the same as having the qualia. It only knows to spit out existing information and recycling or scrambling it according to context via its tensors. Akin to like eating it's own shit. But it doesn't comprehend any of that.

[@sundog.]

@@AnnonymousPrime-ks4uf It's not quite that bad. It can generalize, but poorly, because there is no hierarchy or structure to the information it collects. It's all blindly eaten up and treated with equal value. Humans can naturally determine the value of information we retain, our minds are structured and expertly crafted. A machine can't have a soul, but with enough work, they surely can reason.

[@ussassu]

i don't think this is ultimately the right approach to reasoning, it has to be generalized more be accurate right away, more effective/faster.

[@dapooch2521]

I believe reasoning loops are a big mistake they sound promising right now but it's a big mistake. My reasoning behind this is simple, when you see someone talking to themselves in a recursive loop a recursive reasoning loop we tend to think they're crazy and we tend to stay away from them and these are people there is no way for us to know for sure how quickly the bootstrapping will present diminishing returns

[@JMWolf360]

great video to put people to sleep, couldve been at 2x speed