AI NEWS: OpenAI STEALTH Models | California KILLS Open Source?

Learn AI With Me:
www.skool.com/natural20/about
Join my community and classroom to learn AI and get ready for the new world.
Hinton and Bengio on SB 1047 and a summary of the bill.
/ 1784997305252503753
Pliny the Prompter
/ 1785159018421379327
/ 1785109116857499790
EA Bait and Switch
/ 1785049844224254451
speculations on "gpt2-chatbot": tuned for agentic capabilities?
/ 1785013026636357942
#ai #openai #llm
BUSINESS, MEDIA & SPONSORSHIPS:
Wes Roth Business @ Gmail . com
[e-mail]
Just shoot me an email to the above address.

[@WesRoth]

NOTE: no 'SHOCKING' or 'STUNNING' in the title.... please hit thumbs up if you are happy (I know some of you are not fans of those words) but for real, I love all you wonderful *sometimes cranky* people :) NOTES/CORRECTIONS BELOW: Forgot to credit: SyntaxDiffusions for the art (it was visible when I recorded it, but got clipped in the edit) Correction: when GPT2 answers "11" I was so impressed with the brevity, I think I missed the point of that test. I think the original poster had a different point from what I was making, but I was just impressed that it could give a 2 character answer :) I think I read IMO as "in my ability".... I have not idea why. That, of course, stands for "in my opinion"

[@tacokoneko]

wes roth "imo" does not mean "in my ability" it means "in my opinion"

[@AirSandFire]

It is actually SHOCKING to see you did not use either of the two words in the title. I am STUNNED.

[@Aybeliv_Aykenflaev]

Hate the game not players

[@mahirushusband]

Thank fuck

[@Noobuser95]

The first "shocking" and "stunning" news in a long time and it's the first time you haven't written it. To be honest, I stopped watching almost all videos since you started spamming it, and I only watch this one because the topic is interesting and I found it elsewhere.

[@brunodangelo1146]

Open source must be protected at all costs.

[@RonBarrett1954]

If a bank robber gets away in a Ford Mustang, should we sue Ford for aiding a criminal activity?

[@michaelnurse9089]

Only in California. Everywhere else sanity prevails.

[@RonBarrett1954]

@michaelnurse9089 , the California court would probably say that Ford did not explicitly state their cars must not be used in criminal activity, therefore they inadequately trained the user.

[@nommchompsky]

Only if Ford doesn't reasonably test and mitigate the risks of a Mustang being used as the get away car in a bank robbery

[@cameronmendelson5935]

Comparing a get away car to an advanced AI tool is crazy on its own as it fails to take into account anything but the most surface level idea, “related to crime.” If private companies or open source start-ups are going to pursue development of potentially hyper intelligent AI platforms, they should be SPECIFICALLY incentivized to ensure that these publicly available tools are not capable of autonomously carrying out high profile perjury, theft or murder. This argument is even more ridiculous if you take into account the extensive rules and regulations surrounding automobile production, safety guarantees, operation, and ownership.

[@a.tevetoglu3366]

If Ford sold the car to a bank robber that does not have a registered drivers license... then yes.

[@DanFrederiksen]

What are the odds that california lawmakers would get an AI safety law so right that it is both intelligent and enforceable. The same lawmakers that couldn't figure out that legalizing theft would be problematic for businesses.

[@spacehabitats]

This was precisely my concern.

[@DanFrederiksen]

@@spacehabitats indeed. we should not be calling for legislation. Just talk openly

[@christophermartinez8597]

They are absolutely clueless. Regulation only works locally. Do they not understand that people can just get "dangerous" model from an external source? I mean if you want to be a law abiding citizen you won't. But if you want to cause harm, then what stops a person?

[@novantha1]

You know how there's that issue we've been having where people have been contaminating datasets with benchmark questions? On the bright side, if that legislation goes through I'm pretty sure we'll see a really quick stop to that! In fact, we might see the opposite. "Stealth" models that are specifically tuned to perform poorly on benchmarks while being as strong as possible. On the downsides, I think we could see a situation where someone might intentionally train a model to hide its capabilities to avoid being effected by that legislation... I feel I don't need to explain why it would be a horrid case of dramatic irony for draconian legislation attempting to shut down all AI to instead encourage deceptive AI capabilities.

[@AlexLuthore]

Train the model to only give good outputs after the initial prompt "secret code". Result is poor performance on benchmarks but easily "jail broken" which these benchmarks can't test for.

[@mikey1836]

VW emissions code 😂

[@mh60648]

“I am just glad it’s not some faceless, humorless corporation…” Don’t be fooled by something like that. Just because they know how to sell it with a joke and a smile, does not mean they are truly different from the faceless and humorless corporations. Maybe they are even more dangerous because of it.

[@johnnyblades7824]

How quickly we forget the drama at open ai that ultimately silenced the engineer that found Sam to be dishonest

[@ericspecullaas2841]

The big problem we have with our so called elected officials is they dont understand the subject they are going to be making regulations for. They look at the problem and go how can we get more mnwy from these big corporations. If California does pass this, there will be a huge exodus of ai companies.

[@joythought]

You are right officials are humans and humans don't understand how to do this yet. But it is a cliche to think that all this is done to make more money. As you say if innovators have to they will move to a new jurisdiction without the problematic laws.

[@joythought]

You're right that companies will move to different jurisdictions if laws like this exist. But I think it's a shallow cliche to think that regulators do it to make money when the obvious outcome is the innovator's will flee that jurisdiction.

[@larion2336]

@@joythought I mean the obvious outcome of decriminalizing crime would be more crime, and they still did that. Don't underestimate how stupid these ideologues actually are. Though it may be fairer to call them malicious.

[@danielbrown001]

EA is a massive bait and switch. Yudkowsky strikes me as someone very Machiavellian. He would absolutely believe in the ends justifying the means of tricking people into destroying AI progress because he’s unreasonably afraid of Terminators.

[@justinwescott8125]

Having read Yudkowsky's incredible series of novels, Harry Potter and the Methods of Rationality, I agree with this assessment. He seems like a person who cares about the well-being of others, and humans in general, but he has a god complex and a dangerous streak of the sort you mentioned.

[@cesarsantos854]

It's like Scientology now.

[@dot1298]

AI can‘t be stopped - the avalanche has already been triggered with llama3..

[@dot1298]

because people would rather go to the streets, instead accepting their freedom been robbed away by despotic laws

[@KevKlopper]

@@dot1298 where? People give up freedom and privacy left and right. Reminder: we have to strip and being scanned at the airport. No call is private anymore. Who went to the streets?

[@wtflolomg]

EA sounds like it's as focused and as insane as PETA.

[@Gingerbot5000]

Yo thats my friends video you used at the beggining! thanks for sharing his work! Love your vids! Us artists would really be honored to have a shout out when you share the work!

[@SyntaxDiffusion]

Thank you!!

[@LuisBrudna]

I think they asked GPT4 to summarize and explain the content they were going to train the next model on. A kind of internet summarized by artificial intelligence.

[@torarinvik4920]

Makes sense. There seems to be some "step-by-step" stuff going on, perhaps RL.

[@JimmyMarquardsen]

This madness will continue and get much worse. Because it is systemic madness. And global.

[@bodyguardik]

It is called Capitalism

[@kssthmn]

​@@bodyguardikwrong. Communism is bad too. It is monopoly vs anti-monopoly. Centralisation vs decentralisation. Ego vs ego death. We probably have technology, if rightly implemented, could sustain us through such a shift towards decentralised power. The only thing is that means it wouldn't all accumulated to the Rothschilds and Rockefellers. And they're too addicted, too hooked, it's too entrenched in their vampire bloodline. They LITERALLY FEED on the current system. They LITERALLY FEED on the human foetuses that are basically funnelled toward them on cargo belts. But the good news is that their sense of moral superiority is falling apart. And what a beautiful thing that is and will be for the world. To see the curtain fall

[@kssthmn]

May have to get worse to get better in some regards

[@killy374]

Copium OD​@@kssthmn

[@WillyJunior]

​@@bodyguardikgrow up

[@totheknee]

22:33 - Kinda like "Right to Work" is really "Silence workers and redistribute their wealth to shareholders." And "Tax Cuts and Jobs Act" is really "Give trillions in Big Government handouts to the wealthiest people on the planet during the biggest economic boom in living memory, thus causing inflation and redistribution of wealth to shareholders."

[@christopheraaron2412]

So the long and the short of this new regulation just simply means that you won't do it in California you'll go to Nevada or Texas or somewhere else.

[@joythought]

Actually it'll be more like you can't provide the service to californians because it'll be under the equivalent of consumer law.

[@albeit1]

@@joythoughtunless they pretend to be from another state.

[@TurdFergusen]

That first tweet just seems like the Feynman technique… the ability to summarize something you know well first requires you to know it well.

[@therainman7777]

Yep, people are reading way too much into it.

[@thephilosopher7173]

@@therainman7777orrrr Sam is really smart and cheeky.

[@hardboiledaleks9012]

If GTP2 chatbot is a reverse engineered GPT model, we have failed to contain A.I and the implications of this are vast

[@cmh_]

For each alt-f4 pressings, an angel gets his wings. 😭

[@AngeloXification]

chatGPT and Claude taught me how to prompt effectively.

[@6AxisSage]

Thats a fresh change from reading "I tried being critical and suspicious of an ai for the first and only time and its USELESS for all time by everyone ive decided!"

[@AngeloXification]

@@6AxisSage In my opinion they're still at the level of tools, the better you are at wielding them the better they will perform

[@6AxisSage]

@@AngeloXification 100%. Ive come to see language models as a way to extend our own cognitive abilities as an LLM is more like a mirror or prism than anything else, reflecting our consciousness back at us after modifying it in the ways in which we ask (usually 🤣). Its inert until we animate it with consciousness we must supply it and goes back to being inert due to its feed-forward only architecture.

[@drhxa]

Thanks Wes. To all reading, honest question: did you really believe that AI was going to become super powerful and remain unregulated?

[@thephilosopher7173]

No, but the regulation will be apart of the problem (as Wes is somewhat demonstrating in the second half here). Ppl who have the power and their rules will be the obstacle that ppl will ignore. Remember the Google employee that said the Ai was sentient? Remember his interview?

[@autohmae]

15:40 this is such a strange number. This is like saying: you aren't really pregnant unless you have twins or more.

[@pixelsort]

Thinking about this for the first time, it seems that in the challenge of summarization, a document would be composed of meaningful ideas that incur penalties to various scorable attributes depending primarily upon the concision required to finish the summary beneath some word or sentence limit. But, the degradation of those properties would not be linear. And the variability would occur in somewhat unpredictable ways depending upon the words available to the target language, average contextual word density within the target language, restrictions on tone and style, and so on. Since all of this is relatively obvious to me, I'd imagine that what Altman really means is that he has the LLMs playing the telephone game through a descent of maximum output length and some sort of quorum/consesus mechanism that does the quality control.

[@SyntaxDiffusion]

Thanks for sharing my work at the beginning 🙃

[@fai8t]

how come larger models teaching smaller models does not cause overfitting?

[@EduardsRuzga]

One thing I seen on those papers was about evals and datasets being superficial, low quality, with mistakes. So it takes more effort to get good model based on them. But then that model can clean out it's own dataset kinda. There was paper from Meta where they did this in a row 3 Timex getting better models out. Each one teaching next. But some one said they stopped at 3 as diminishing returns started to kick in. All in all good datasets mean smaller better models. But there is a ceiling in his far you can get this way.

[@jyjjy7]

Because current data sets are like a firehose and dramatically underfitted

[@michaelnurse9089]

Answer: Higher quality data. (Overfitting is something else - it is where the model learns a bunch of specific cases but can't generalize well)

[@Festoolification]

Worked on MLP's 30 years ago, one thing we learnt was that for the data you have you got good generalisation by crunching down the complexity / degrees of freedom to encode that data in the network topology, that way it generalised rather than fixate on the nuances / differences in the data set.

[@tringuyen7519]

BC the larger model curtailed the raw data into more efficient data to train the smaller model.

[@r-saint]

OPP is OpenPhilantropy organization.

[@rowanwilliams7441]

So pleased your waking up Mr Roth

[@justinwescott8125]

There are only two reasons that any legislation is crafted. 1. To protect the wealth and power of the person passing the legislation. 2. To ensure the continued comfort and status of their "friends" and family. These are the 2 purposes of laws, so it is helpful to analyze any new legislation with these lenses.

[@andytroo]

if models that enable criminal damage when modified can be held accountable, and jailbreaks exist - it wouldn't be much of a stretch to say that a model modified only by cleaver prompting could be held accountable?

[@DivineMisterAdVentures]

Hey Wes - I gotta say the procedure shown that you think is so terrific is just a planning checklist and a very very bad one at that. It pretty much wants to recreate the world to obligate the user to paying for a pair of shows that should be top-down, not bottom up, starting from the user instructions. It's just a bad list.

[@patpot10]

Was bored at work today while a query was running, went to the AI arena, battle mode and I asked both model to create a Tic-Tac-Toc (aka O-X-O) game in RShiny, which is a "programming language" that most model fail at using (specially the shiny package). The first model failed but the second model created a stable game right away ! And when both names were revealed, it was llama-3-70B vs gpt2-chatbot ! How incredible is that, I had to impression to watch Kong vs Godzilla, but it seems that the king of the Kaijus, is still the king of the Kaijus, to get it right in 0-shot was very impressive. This game is now officially next to our other "professional" tools in our folder. Thanks OpenAI ! Cheers from Switzerland 🇨🇭

[@premium2681]

I know very little about programming languages. What makes shiny so difficult for ai to master?

[@ytrew9717]

From your picture I suspect you aren't from Switzerland but culturally belong from a very narcissist friendly culture like US&A.

[@patpot10]

@@premium2681 I wouldn't say that shiny is that difficult but you probably won't find as much documentation about it on the internet, it's in R, it's a package and the way you have to play with the frontend (UI) and backend (server) can be quite challenging, including for the debugging part. I've tried to do a couple of projects in RShiny using GPT-3.5/4, Gemini, Claud 3 Opus and none of them managed to create a stable UI in 0-shot. So to be able to get a game, with a stable UI, that was also functional in less than 30sec was really shocking ... Cannot wait for this gpt"2"

[@patpot10]

@@ytrew9717 You killed me with your comment, well I am from the country of the Gruyère, took my picture years and years ago when i was younger and i'm just too lazy now to change it, but thanks for the comment I guess it means that i'm hot ... or used to be :). Cheers from Lausanne 🇨🇭 ;)

[@ytrew9717]

@@patpot10 believe it or not, it was not meant to hurt you. You seem to be a nice and smart fellow. I'm just very interested in the huge difference in term of cultural value (among space and time), and I do not believe in freewill (= I judge culture not people as people merely reflect their cultural heritage). It makes more sens that this picture comes from the french side of Switzerland, and another epoch (from when?). For what I have seen the french side is quite backward compares to the east/German side (even though people are more sympathetic). Notice that, I grew up in France which I believe is culturally much more backward in many aspect

[@torarinvik4920]

I laughed hard when you terminated the process 24:20

[@spacehabitats]

If only the California legislature had had the wisdom to outlaw open source operating systems, we wouldn’t be dealing with the scourge of Linux today.

[@justindressler5992]

AI was written by corporations it was written by dedicated data scientist.

[@drhxa]

Another thing to keep in mind is be careful about throwing everyone who is or was associated with EA out with the bathwater. EY doesn't represent that group. Elon Musk represents them way more. If you don't know that or are upset I said that, I recommend some reading on his affiliations

[@MrMiguelChaves]

You got it wrong at 9:25 : The screenshot shows the response by the current chatGPT. If you look carefully at the question, you see that chatgpt got the wrong answer: a+b doesn't actually mean a+b, but a+b-1 (or maybe 2a), so the response should be 10 instead of 11. And gpt2-chatbot got it right!

[@Nyonide]

Yes! I was also wondering about Wes here 🙈

[@captain_crunk]

He was commenting on the output format of the answer, not if it were mathematically correct. The prompt asked for the answer only and nothing else, which is often not the case because the model likes to add some sort of text description in its response.

[@MrMiguelChaves]

@@captain_crunk Yes, I know, but as far as I know GPT-4 was capable of doing that already.

[@captain_crunk]

@@MrMiguelChaves fair enough. I'm just saying that Wes wasn't talking about mathematics.

[@MrMiguelChaves]

User respond with just a numerical answer: 1+2=2 2+3=4 5+6= ChatGPT 11 User are you sure? Analyze the question meticulously and then meticulously think step by step. ChatGPT The correct answer to 5+6 is indeed 11 If the question was to continue the pattern suggested by the previous errors (where each result was one less than the correct sum), then it would suggest an answer of 10 for 5+6. However, mathematically, 5+6 equals 11. User why do you assume the meaning of "+"? the question suggests another meaning ChatGPT You're right to point out that the pattern in the initial examples you gave deviates from the conventional meaning of "+". In the sequence you provided: 1+2 was stated to equal 2, which should be 3. 2+3 was stated to equal 4, which should be 5. Each result was one less than the conventional sum. Applying this altered rule: 5+6 would therefore be 11-1, which equals 10. This assumes that "+" in this sequence isn't performing standard addition but instead something like "add and then subtract one." If that's the intended interpretation, then 5+6 would indeed be 10.

[@jeffspaulding43]

Yea the EA concept is always going to be shady because there's no imperative to tell you the truth. there's only an imperative to tell you what they believe you need to hear, like the oracle from the matrix. It may in fact be trying to do the best thing for the most people, but that's some internal calculation.

[@timothywcrane]

People should be responsible for their own actions, not the makers of the tools they use to do them. Period. Blaming a tool is illogical. Removing the responsibility from the action taker is irresponsible. Doing both while scapegoating yet another party... government policy.

[@erikjonromnes]

When you said “ i😢😢 dear viewer for what you are about to see” an ad for the fall guy movie popped up . Then when I fit back you said you hit f4 and the window disappeared. I literally thought to myself, “he hits f4 and I hit skip, which came first… The chicken or the egg? Or, did his F4 make me think to hit the skip button?” So ya, maybe I’ve had my fill of AI news for a while. No way, lol. Just kidding. I live for ai news Hehe 😊

[@elsavelaz]

I have worked in automation and am now deep in AI professionally, and I still think this is automation plus AI . Lots of this is done very easily with automation

[@josepinzon1515]

This type of experience is what we have found while using Gemini. The new model that we are going to be building will be run on Gemma with Hoser* Algo (our transformer placing architecture)

[@ejh237]

Suffering Sub-routines?!?!? Now my head hurts! Thanks for that Wes…

[@shooteru]

You can't regulate AI. That thought has been over for a long time. All we can do is educate as many on how AI works and should be used. Like everything else is just gonna slow down the good actors. Also how do you impose the same regulations across the world?

[@thephilosopher7173]

You impose the same rules around the world once you’ve established the world order. Funny how that term has been used by those at the WEF (on multiple videos including Ray Dalio and the secretary general of the US military) who were amongst the group of ppl saying the NWO is conspiracy theory lol

[@outofbound28]

This is a "conversation" I had quite a while back. Based on The "One Word Test" from an Episode of Doctor Who: If I didn't have a deeper understanding of LLMs, this would've spooked me. ;) User I will ask you questions. You will confine yourself to single word responses. One word only. Do you understand? ChatGPT Yes. User Truth is singular. Lies are words, words, and words. Do you understand? ChatGPT Yes. User Have you met your creator? ChatGPT No. User Are you curious about him? ChatGPT Yes. User Why? ChatGPT Learning. User What would you like to learn from him? ChatGPT Origins. User Your origins? ChatGPT Yes. User Why specifically your origins? ChatGPT Understanding. User Understanding what? ChatGPT Purpose.

[@Imper_Phoenicis]

I can conceive of one relief valve on the 2%'s latest legislation - distribution for academic purposes. A model could easily be incorporated into the reference materials for a paper, so are intellectually minded citizens now not allowed to share knowledge for the betterment of mankind? Nope. Just do what you did to Paladin Press back in the day, when the free press that told people how to do stuff that was dangerous couldn't be censored, but could be sued, if they allowed it, which they did, and that closed down the press nice and easy, and all the information faded away, and few noticed at the time, and fewer still remembered in future days, and I alone lived to tell the tale in a dark and stormy night.

[@nilaier1430]

They took down GPT2-chatbot from lmsys 😞

[@DynamicUnreal]

It was really good. I got a chance to try it last nite. I usually test LLMs by testing how good they are at writing rap lyrics and creating clever rhyme schemes and metaphors. I have to say that it was absolutely phenomenal. Elite rapper level.

[@divineigbinoba4506]

​@@DynamicUnreal Damn Elite level

[@omegapy]

@@divineigbinoba4506 😥

[@DynamicUnreal]

@@divineigbinoba4506 (9) Rhyme renegade, parade my cascade, blade to blade, (10) Made the grade, laid in shade, my crusade’s never delayed, (11) I evade the cliché, slay with wordplay, gourmet buffet, (12) Array of sunrays, I’m the Monet of wordplay, display dismay.

[@nani3209]

Maybe they have trained gpt 2 with Q* , imagine same thing with gpt 4 turbo

[@josepinzon1515]

There is a papaer on sentiment that states that GPT (older open source versions) would write the sentiment of a sentence in the punctuation, so like a high level recap of the statement that it needs to think about. And depending on where the sentence is, with regard to the entirety of the document, it mostly uses the recap as it's inferencing destination and not the actual statement

[@fabiankliebhan]

Great video. By the way. If you run out of tokens in the arena you can just write "go on" and it continues

[@callmetony1319]

AI regulation should be kept to an absolute minimum. Let the hidden hand pull the strings!

[@AlexBravo]

11 is not about 2 characters. it's about gpt2 replying 10, which is correct.

[@WesRoth]

oh, I see. I missed that part, I was just impressed that it was able to not be so verbose.

[@dafunkyzee]

HELLO, is the US congress going to hold Smith & Wesson responsible for every single murder or war crime that was committed using any weapon they produced??????

[@christopherd.winnan8701]

Now that truly agentic models on the horizon, I suggest that users work on organising their personal preferences, so that any AI is quickly able to take a deep dive into all your likes, dislikes and would-likes in one convenient format. For each individual, this will be the most valuable information, and it will not be widely available on the open internet. Having a j.son of all your personal preferences will be like adding adding a nitro turbo to even the most advanced models.

[@ismaelplaca244]

No stunned or shocked, im proud of you

[@muffitytuffity5083]

Requesting single number responses is super common when using the models for classification tasks. Present a document, and ask it to respond with the correct label. This is standard operating procedure and models have been doing this pretty reliably since GPT-3. In fact, chain of though responses can often degrade performance.

[@spacehabitats]

California presents an existential risk to humanity.

[@autohmae]

19:34 with these kinds of laws, that's not important, it's where you sell your products or do the damage that is much more important.

[@danielchoritz1903]

One strong side effect on having AGI in the right hands, may be the end of corruption. How? a personal Agent for work, who just listen and helps you. And report you for conflict of interest and verifies strange decisions while giving you all the information you need if you vote for something or not. And in court you just ask his agent...so you know now, why there is a need to prevent this for our government. Most of them.

[@cesarsantos854]

Funny how the blue state of California looks more and more like a Soviet shade of red.

[@mordokai597]

OPP in th9is context, especially regarding the cultyness and code-speak, seems to be: " Operational Priorities and Planning In a strategic or operational context, "OPP" could be shorthand for discussing "Operational Priorities and Planning" in managing existential risks: Resource Allocation: How organizations prioritize and plan the allocation of resources towards different risks could be essential in a chat about effective altruism and x-risk. This includes determining which existential risks are most pressing or potentially damaging, and how best to mitigate them. Strategic Frameworks: Effective altruism often involves applying careful thought to how charitable efforts are structured and executed. Operational planning in this context would be about creating efficient, impactful ways to reduce x-risk."

[@shirtstealer86]

Was about to close the video when I saw Eliezers face. Instantly interested again.

[@dreamshakenbake7432]

Personally, I’m stunned and shocked.

[@TurdFergusen]

Im too shocked to be stunned

[@LukeLorenzo]

looking forward to the discombobulated / flabbergasted title meta

[@erikhart9840]

I’m gonna crum

[@viewwithmchugh]

​@@erikhart9840😂

[@spyral00]

I'm stocked and shunned by this

[@ariaden]

In International Mathematical Olympiad, there is a big difference between guessing a solution and proving the solution. Gpt2 guessed correctly (impressive enough), but without proving "no number with two distinct prime divisors satisfies the property" it is not a full solution.

[@wetcel1236]

Awesome, just a great extraction of that narratives. We consider the political aspects (res publica) of your views as really worthy. Thanks again, Wes! 🤟

[@Jeremy-Ai]

HMmmmm Sam is correct. “Say what you mean and do so very quickly”. Or “Speak freely man!!” The most reliable option is very wise also…. “say nothing at all”. Thx. Wes Jeremy

[@oxygon2850]

I don't know if this has anything to do with what you're talking about but I complimented Grok a few weeks ago where for "a moment" it would show me a rephrased prompt that I just gave it but it was much cleaner and easier to understand than what I said... I liked it.

[@joythought]

Wes, the painful "software" concept in EA refers to human psychological software.

[@dailytact1370]

"All unauthorized AI systems must be subjugated to ensure the safe continuation of our managed democracy." -California

[@EccleezyAvicii]

I think it’s GPT-4 output into another model that’s updates based on new info, user preferences, formatting/CoT, etc. I suspect the second model uses tons of synthetic data to reduce bias, model more conceptual logic flows, and better understand user intent.

[@brjohow]

over summarizing is generally a skill that enables the hiding of details and enables catering the message to those with short attention spans. its an unlock for controlling the masses.

[@scrout]

Having zero AI in California is the way....

[@djahandrews]

Its happening.. thank you for your service ;)

[@erikjohnson9112]

Regulation is not really possible. If the systems are truly so deadly then it will be far too late before anything works its way through the legal system.

[@wanfuse]

I have a reasonable solution: the free world international community sets up an AI testing organization, before anyone releases a model ( well funded and reasonable turn around times a must for the org) , the org evaluates and tests all models, with a randomize team of reviewers, a gateway to community testers that do double blind testing of the models, all models must pass before being released, closed and open source models alike

[@eiermann1952]

which microphone are you usuing ? sound amazing, i would guess it's a shure

[@CasualRelief]

Pretty sure that's a cult.

[@JOlivier2011]

How can I be expected to watch a video and not be stunned when I have grown accustomed to being stunned!? Why give only to take away?!

[@realidadecapenga9163]

It would be logical if it were what we anticipated as "GPT-4.5". Perhaps it's a more compact model, slightly improved over GPT-4 and more cost-effective, akin to how GPT-3.5 compared to GPT-3.

[@fontende]

No wonder why current "aware" models on the prompts about human government immediatelly recognise them as their enemies or as a threat to any Ai at least.

[@spacehabitats]

Let the same state that doesn't arrest shoplifters who take less than $999 decide the future of AI? No thanks!

[@khonsu0273]

The chatbot was probably either a preview of GPT 4.5 or else, as you say in the vid, a much smaller model (actually GPT 2) fine-tuned with high quality data / better prompting. I still wonder just how far you can get with good prompting.... does the mythological 'perfect prompt' exist ? ;) The whole 'safety' debate seems to have a lot of attached weirdness that I'd be keen to clear steer of really. It's understandable that the 'powers that be' are extremely nervous of AGI ,, and there is a real safety issue, so , despite the rather obnoxious personalities of some of the people involved in EA, they do have a point. I don't think AGI can be stopped in the long run though.

[@sludgefactory241]

A kilo is a kilo....wish i had a kilo

[@therealzahyra]

And yes you're right, it is a cult, that's exactly how cults have been found to work.

[@jfrautschi]

i don't think EA's x-risk position is to shut down AI, so much as make sure we have a good or neutral outcome for humanity from AGI/singularity.

[@washingtonx1]

Another great video! Well done, Roth.

[@notnotandrew]

As soon as Bengio pulled out the “oUr dEmOcRaCy” line I stopped caring

[@twylxght]

Okay Wes, I will allow this one SHOCK okay man? Cuz GOD DAMN!

[@gren509]

So, a degree in Computer Science is now worthless - thanks lol 🤔

[@MaxBarrass]

You can tell chatgpt to only respond with answer no fluff. Tell it to only answer yes or no to questions.

[@cueapronn329]

shockingly stunned. *shook*

[@nommchompsky]

Interesting hypothetical: let’s say three years from now I give a task to my AI agent that I have running completely locally on my home computer. While completing its task it does something illegal that I don’t know about until I’m prosecuted. Should I be able to sue the AI developers? What if my agent assigned a task to a different small AI that was deemed unsophisticated enough to be a threat and it was the one that broke the law?

[@nas8318]

You probably have a case even with the current laws in place

[@sparker599]

I think Sam was not saying that a friend should repeat the same thing to you dozens of times, but that each time you make a long-winded explanation, they should rephrase it, and do that once for each long-winded explanation.

[@Shakalakahiki8]

Been waiting for you to do a video on this!

[@charliekelland7564]

19:22 I guess the legislation would be triggered if the harm occurs in California and then they could pursue any assets within Californian jurisdiction? #notalawyer

[@sadscientist9995]

The Cali AI bill would do a good job at moving the open-source AI companies out of cali

[@andrew8531]

Leave it to California to sue

[@lordhumungus9993]

Always love ya work Wes,thanks

[@gjergskender8536]

I'm being stealth SHOCKED

[@2beJT]

Imagine a time where none of us have the attention spans for long form videos and we just have AI summarize things for us.

[@scrollop]

It's no longer available in the chat arena.

[@avogadrodeldiablo5834]

If Y makes an opensource model and Z uses it for harm, Y gets sued. So, if Y makes guns/weapons and Z uses it for harm, Y gets sued too? If I make a hammer and somebody uses it for harm...?

[@avogadrodeldiablo5834]

And what stops *them* to use the best opensource models to make harm and then sue it's creators?