AI News: OpenAI is up to something *new*
VloĆŸit
- Äas pĆidĂĄn 24. 07. 2024
- Rumors of OpenAI's next big thing as they block China and announce a new "app".
My Links đ
âĄïž Subscribe: / @wesroth
âĄïž Twitter: x.com/WesRothMoney
âĄïž AI Newsletter: natural20.beehiiv.com/subscribe
#ai #openai #llm
STORY LINKS:
OpenAIâs China Ban Doesnât Apply to Microsoftâs Azure China
www.theinformation.com/articl...
OpenAI blocking API access (incl. API) from China since today
pro.x.com/i/decks/17310333028...
Factorio + LLMs
x.com/naklecha/status/1810385...
Noam Brown @polynoamial
x.com/polynoamial/status/1810...
AI-Driven Behavior Change Could Transform Health Care
time.com/6994739/ai-behavior-...
OpenAIâs China Ban Doesnât Apply to Microsoftâs Azure China
www.theinformation.com/articl...
Chinese developers scramble as OpenAI blocks access in China
www.theguardian.com/world/art...
CICERO
ai.meta.com/research/cicero/h...
00:00 OpenAI Blocks China
01:55 New Release Rumors
02:35 AI Coach AI, backed by OpenAI
04:05 My Thoughts on "AI Health Coaches"
06:44 Why is OpenAI so quiet?
12:45 Noam Brown
The China "equitable" argument for access is a red herring for western journalists, for sure.
And also the developers in China, like everywhere else, just want to get to work. Sure, as in any group, there are a nefarious few - ones who share much in common with their western counterparts who occasionally take a break from spying on their own citizens to engage in anti human activities. The vast majority simply want to build without virtue signaling political interference from self important flatulence.
1,000,000% you would have to be a fool not to see this for what it is
Can you elaborate? Do you mean they're trying to bait western journos into siding with them?
â@@Penrose707most can't because all the information we consume has very strict guidelines around delivering that message. As soon as he started delivering that message, all of his views would drop to nothing. I can prove this with two different Raekwon the chef videos about him singing about Osama bin Laden lol. The first one 6,000 views in 2 days.
The second one on a similar tiny channel? Like 30 views lol.
I think that automating video games is a huge step towards generally useful AI models. There's so many things you have to get right in order to play a game, particularly an open ended game, and plenty of those things have crossover with the real world in a variety of capacities. From things like object permanence to strategizing, to operating in an environment in an agentic way... These are all things that somebody is going to randomly figure out one day and go "oh, that's neat, I can automate X task I don't want to do in the game" only to fail to realize that "Oh, there's a related task in the real world that doesn't need to be done by people anymore".
Yes and generative AI merging with VR & mixed reality is gonna be huge.However that requires insane processing and energy but I can be done
Oh absolutely calling BS on any chinese model surpassing OpenAI
as someone with limited use of their hands, I LOVE AUTOMATED GAMES! when you're mindful of how much you can click, the endless repetition of idiotic game loops gets really frustrating, I'd much rather sit back and make the meaningful choices
bruh wht, the loops are good. we can respawn
9:05 "chatgpt" has become a household name like "kleenex" at this point
Or Google as a verb despite other search engines
@@Edbrad exactly
AI + Nudge = Big Brother 2.0
Google had tons of political power from their rank ordering of searches.
Imagine the power to control a person's reality when you're *composing* the answers to all the questions a person is interested in.
It's Big Brother 3.0 when AI goes from oracle to coach.
Finally someone's thinking đ
Suggestions to be healthier and more productive? Oh the horror! Just like any other person that has fear mongered technology advancement, you will be left behind.
@@jmd489
PhD in Electrical Engineering
Unlikely I'm being left behind technologically.
More likely I'm better able to navigate between beneficial implementations and harmful ones.
If you're not wary of the harms tech can impose, you will not be as successful.
Keep in mind that the chart shown does not include app usage statistics....
This is a very important caveat I was not thinking about when looking at the chart. We don't know to what extent those logins are single uses versus extended sessions. It is entirely plausible that perplexity and claude have a much higher average use per login than ChatGPT (I'm sure it's not enough to make up the difference, but still...)
Terrifying privacy issues
@ Wes Roth what do you recommend for merging models together? For example: 5- models that are all around 5gb in size. 7B and 4bit-Q.
Bro you realize he is the furthest thing from a researcher đ? I got you though. Google âevolutionary optimization of model merging recipesâ. One of the original authors of the original transformer paper is behind this method. So you already know itâs đ„. The best model merging method Iâve used. If youâre really built like that. The most effective merging method(applies to fine tuning too), use your downstream dataset to determine subnetworks with the most semantically relevant representations(this is extremely useful since itâs hard to know what knowledge models lack), freeze the rest, this is a pathway to fine tuning with no forgetting too, you can use this to create non-redundant new subnetworks, knowledge sharing from prior subnetworks. Working with subnetwork vs whole dense models is superior in every way but way harder to implement.
Good luck in your journey.
The idea of a health coach is an interesting idea... but also imagine a personal assistant that is mostly helpful... but also designed to guide you in certain ways... just in minor ways that gradually guide you to the desired destination. For example, it guides to you do something just before lunch time, so that you end up buying lunch at a particular establishment. Or a model could be built to give different answers for a particular person - encouraging them to do risk taking behavior for example.
I want to see Ai play Age of Empires 4 and what crazy tactics they come up with
The benchmarks were never designed to be a proxy for the typical tasks that LLMs are used for today.
What are?
Can't wait till my AI doctor advises me to eat three rocks every day...
Don't be afraid... it's going to be OK. đ
It could be worse, they could tell you to eat three seashells.
I can easily imagine a SF movie with a worldwide BlackBox AI trained by that only one company which will recommend (if not force) you to take that medicine/injection sold by a sister company *wink*wink*. I will always prefer many diagnostics over 1 unique one, we are not ready for AI yet in our society, we need first to get rid of the concept of "money over everything" imo, then...AI can come.
Donât forget the glue
a few pebbles a day never harmed anyone
Automating complex games is fascinating because it is a herald to real industrial automation
Itâs like developers canât use proxies đ they are not tech savvy enough
Your camera is overexposed and your mic is too hot (gain too high / distorting)
I've been using gpt as health coach for about a year now. When memory became avaliable it went from being extremely useful to extremely useful and effective.
We just have to wait the inevitable required time for them to train their new bigger model at the real next scale. Historically it's been 2-3 years in between each, I don't understand why so many people get all shambled up about it not being here right now. If it gets released this year that would already be much earlier than we should normally expect.
I tend to think that the release time will continue to get shorter as the technology itself plays a larger part in improving itself.
I used to multibox 4 WoW accounts on an RP-PvP server, and then I went on to play a variety of technical roles at LARPs and stuff. This explains, in part, why my AI work is so weird and nonstandard.
Interactive experiences like LARP relates to the stuff like healthcare bots in ways we haven't clued into yet.
I guess I'll pivot the Hipster Energy Club into some sort of AI integrated philosophy and ethics as a service platform.
Hi, probably gpt 4o api multimodal prompts and embedding. your videos are very informative and helpful. thank you. AI news is addictive.
I am making my own prediction of OpenAI release date. Nov 30, 2024. Why? This is the same date that ChatGPT was released which was Nov 30th, 2022. This would be a smart move for the history books. Nov 30th, will always be remembered as the day that OpenAI changes the world. Since they are doing 2 year cycles, GPT6 would be Nov 30, 2026 which would align toward possible AGI like systems in 2027.
Also importantly after the US general election on Nov 5th.
âEat 3 rocks a dayâ - new meaning for âGoogle Docsâ đ
Some day ai won't be that dumb and the people who hang on to the little mistakes it makes in it's infancy will be left behind. Same group as trump tards.
AI personalization is going to be the name of the game going forward. Whoever makes the AI that will seamlessly plug into your life and be personalized to you will win.
OpenAI has the brand-product association
I really like Gemini...
Great insights on OpenAI's collaborations with health startups and Hollywood studios. These partnerships could lead to significant advancements in both healthcare and entertainment industries. Looking forward to seeing what comes next!
Thanks for the great coverage Wes
I am sure this has nothing to do with the new board member.
I've asked various models for information on health. It sort of gets things right, with more reliability it could be excellent. I also had it create a meal plan, based on macros, for each of the days of the week based on what my exercise output is and then asked it to make a shopping list. It worked pretty well tbh.
Claude is my goto for accuracy at this time but I periodically hit 3 with the same question and see which is currently giving the best most accurate answer.
Trackers are good but what I found a few years back that many problems like calorie counting on pool laps or pull ups had no accurate app out yet. With AI now I think these are provable right or wrong but since undone yet there was no measure of truth. It is happening now with people solving base truth through consensus, fact check and reference. Great video Wes!
does the Ai share data with the developers? because i spent the last month talking with the AI about creating personalized AI
Apple in its developer showcase a few weeks back said that Apple will never have access to your data as the AI model will encrypt data and will utilize encryption when offloading processing to Apple's AI servers. so in their own words they claim no, but they didn't explicitly state that OpenAI wouldn't have access to data, just that they(Apple) wouldn't.
logically thinking about things, I believe OpenAI will have access to data in some form to help train its models and help evolve them. to me that makes the most logical sense, but again - Apple has explicitly stated that they(Apple) won't have access to your data.
I was tracking calories for some time and it really helped me losing weight. But it was a hassle entering all the ingredients of my food, especially if you dont cook it yourself. I imagine an ai tool as a personal health coach to be the most useful thing for me in the short term. Imagine you just taking a picture of your meal and the ai estimating the calories and nutrients with an error bar, so you can take the error into account.
*Reads title*
*Has internal monologue* ...Oh boy.
True AI Assistant would be awesome. Speech mode also very useful seems. Helping to connect and be mentally healthy sounds great
On the growth stats: Those are just the *_web visits_* - it doesnât count mobile or desktop apps, because those are connecting via API calls. (Their compute bill for inference must be staggering đ€Ż)
So what happened in April 2024 Iâm guessing the announcement of the stuff we donât have yet
i agree with voice being better for frictionless computer use for things like giving us realtime info on our life, the issue is who gets access to this data, security, altho it would be good still but the corps would get too much access. we need a open local solution for this
TENCENT steal their code and misuse it? lol
8 reps, 40 lbs? You a school girl, @Wes
I hope they will release the older model weights if a new gpt5 or w/e comes out. We still cant even run gpt 3.5 local or really mess with it. What happened to the "open" in open AI? đą
My gut says GPT5 would and should hold off until after the US election.
On the other hand, it would be nice for voters to know before the election what the president will have to deal with in a post GPT5 release world.
So OpenAI isn't open in terms of open source, but that's okay because it's open for all to use to benefit all humanity (China not included)
đđđ
I would love to see more content on how to use AI in regards to games/making bots or companions (skyrim). I'm not a developer but I love technology/video games and learning new ways to learn :)
9:26 OpenAi have so far has become the face of AI, the same way Google became a verb, despite other search engines.
Very often I wonder why I still need to manually enter my food into Cronometer.
90% of my consumption is repetitive behavior, depending on if I work from home or the office, if family comes to visit and especially on what I bought before.
I would be really happy about AI support in that field.
if openAI had something new/revolutionary AI wise and asked that new state of the art AI thing to ruminate about the best strategies going forward, it likely would offer strategic possibilities so new and unique as to be beyond anticipation for to those behind the curve....
. Will it eventually control you through the access it has. Are there parameters it won't cross
We didn't even sign up for this
@@joshdickey2805 you don't have to buy an iphone or a macbook or any other product. when you buy any company's product, you quite literally sign up for whatever they have planned via Terms and Conditions agreement.
Open AI : a lot of announcements, but nothing more as usual. It's always "soon"đ€
They still haven't delivered on the last "new" thing. 4o roll out isn't complete. No GPT 5. No Q*.
i was playing factorio while watching this lol
You're excited about LLMs being able to play games because it means that they are able to adapt to an environment and take action. They have some amount of agency, predicated on their ability to understand some end goal and coming up with and executing a series of steps working towards that goal. That's how the real world works, in my view. It seems like if you get these things intelligent enough then they'll be able to play warehouse just as well as they could play factorio - just a bunch of cameras everywhere and it makes decisions on what every robot should be doing at any given time as if it was playing a game.... cause it is. Sometimes I think this stuff's super close, and other times I feel like I'm being duped - only time will tell but this roller coaster ain't the best for my sanity!
Behavior change through micro decisions guided by the gpt4o combined with meta type glasses. Ive had this planned over 6 months now.
It's been known for some time that well tuned models can be used to jump-start the training of new models - that is why China is being blocked.
Wes, I love your content and have huge respect for your precious input. The only thing I wish you did was to make sure every slide you are presenting is in night mode (much like Twitter which you thankfully do). My eyes are not as young as they used to be and staring at a white graph for 5 minutes straight is no longer a task I'm up for. Cheers.
AI for everyone, they said earlier. for the good of all mankind, they said
I'm a little surprised by how niche claude still appears to be
Me: I felt so good last week what was I eating
(confabulating) AI: dog food
WOO!! Awesome stuff as always, Wes! :)
More users, GPT's intelligence evolves faster.
Bold assuming they're gonna release anything
ChatGPT wouldn't exist without magnitudes of out sourced data set. Why would training from ChatGPT be different?
Always something under radar then woop pop out
Wes youâre thinking of âLocalizationâ,
its like a âcultural translationâ for slang terms and the like,
I have a fully fleshed out 4 player board game that I play regularly with engineers for ten years. I approached a bunch of devs and send dozens of letters - canât get a reply. Thereâs nothing like it but just a itâs on my shelf. I hope AI gets good enough so I can make it myself. Be no harder than making online chess (which is still too hard for me personally)
The hype train needs more fuel!!
I predict a big reveal aligned to an Apple product announcement, so maybe around September to support the iPhone 16 hype train following their recent elevation to the board. Of course, doesn't mean there wont be another one as well.
"equitable" is actually a thing in China's governance. They actually practice better capitalism than the US because of this, as they don't allow any corporations to monopolize everything. I know, it is shocking to the increasingly far-right governments of the North Atlantic countries. But the US used to be like this too, a very long time ago.
Itâs almost like OpenAI asked ChatGPT how to make a growth strategy to dominate the market, right?
You're talking like they will release gpt 5 but they still haven't release 4o yet..
Securing against potential cyber risks are probably actions recommended by the new board member who is a retired Army Cybersecurity expert.
The problem with AI giving health advice is the same as a human doctor/nutritionist. Garbage in. Garbage out. So much of our current health advice comes from deeply flawed observational studies, often based on food frequency questionnaires. If the data the models is exposed to is of low quality, the advice it gives, however individually tailored will be suspect.
BTW, there's already a game like Risk where you have to negotiate alliances. Diplomacy.
What's up Wes let me get a comment. I'm your #1 fan, this is Stan.
You killed your wife and kids on the freeway
â@@efraimmukendi7137I remember that. Horrible.
Remember nwa? Momâs spaghetti?
AI mastering games like CoD and WOW is very exciting for me given it's human dynamics (novel, spontaneous, cooperative, etc). As an ex-competitive-WoW healer, if an AI can perform that at a top level there in a raid, it can do almost anything in the real world. I dont play FPS games but imagine what level of fighter it would be if it was the best in Call of Duty? Scary.
Lol that's not gonna matter in competitive tho, it'll be more like a cheat. If you wanna kick back and enjoy hacks/cheats sure that'll be fun, but at the highest levels they're not gonna allow it, it's still gonna be human only.
But yea AI, once really implemented in the real world is gonna be groundbreaking, it really will be a "do anything" kind of ability. It's exciting to me.
16:33 could be worse, you could have been hyping your audience for months that we'd have AGI by September lol
The problem with AI is that it's energy inefficient and untrustworthy and it will clash with the green zealots! The only good thing is that might clean both and pop the AI bubble!
Canât wait until my AI health coach tells me I need more PepsiCo convenience food items, but should avoid at all costs General Mills and Kelloggâs if I want to be my healthiest and happiest self, and to relieve my stress I should immediately purchase a variety of feel good fidget trinkets from Temu (Iâll go ahead and order it for you!!)
All with superhuman persuasive capabilities! Real time behavior nudging? What a recipe for disasterâŠ
So we never getting GPTO voice ...
Minecraft Baritone
NSA at work... SMH
Chatgpt5 wen?
November
December â@@user-ty9ho4ct4k
Definitely before GTA6
A rotten state with Ai in his hands...
How about US hospitals actually give data to AI so it can help in many ways and not restrict it like they did with Watson. I mean they might lose billions but people's health is important right?
great
I go back and forth between ChatGPT and Claude every month or so; i find claude to have such a better personality.
How would the average human score on AI Benchmarks and tests?
I won't be impressed by AI until they figure out how to use it to solve the housing problem.
You mean even more announced but unavailable features.... I'm already containing my lack of excitement.
Not all can use cloude3
how do i contact sam almant about my long term project ?
â€
I won't be impressed till AGI.
@@anta-zj3bw dude. Natural language computer interface with significant agentic capabilities will be amazing.
Well they gotta try to present they are taking precautions based on former mishaps delayz and being in news not so positively gotta if wanna remain world leading dominate and convince and sign big collabs have right companies behind them before release to public sell selves makes sense
3:30 oh no my ai phone is going to start nagging like a woman..
It's socioeconomic pressures that are causing poor heath choices. You cannot address the necessary changes in people unless their environment changes. Your perception of the success of these hyper personal health guidance models is from a privileged perspective.
So let me guess.
It starts with A is for Apple, M is for Meta and Military & N for NSA chef both on the board add up to...
Did you see Claudeâs new âpublishâ button with Artifacts?
I stopped caring about openAI and doubt it has anything left in it. Sutskever was the inventive brain. (There are many PhDs, but only a few are geniuses).
GPT5 is definitely around the corner, plus at the end of the day we don't know squat of what kind of minds they have left, who else is joining etc
@@ToxicSapiens yes but he gave them the blueprint for AGI before he left. That's why is is starting the SSI company. He gave them the super villain, now he is building a super hero.
Wes, I think it's pretty obvious that OPenAI is being nationalized eg Los Alamos, hence the timing on their position with China and opening up use of the model to the world no registering required, aka honeypotting it. All your prompts are belong to PRISM.
So much r & d taxation falls on these models & tech. I mean in the entire transitor age and before .
The ability to grant Americans access to become owner's and operaters of skills and trade or sub contract themselves out as such is at a critical extreme levels of need.
Idk how usa can do this as is status qoue. Even our allys undermine our ability to become kings of our own castles master of our own domain to sense the universe in all its glory. This wasnt just a slogan .
Accessible simulation power for private scientist to manipulate molecular structures cant just be kept under lock n key forever and it cant be given out to the world .
Millions of American native family Self-sacrifice has ran to deep to long .