PROFOUND Step Toward AGI--DeepSeek R1 Explained

[@HarshColby]

It's good to see researchers focusing on having models learn simpler concepts, then using that as a base to learn more complex topics. Seems obvious this would result in smaller and better models.

[@GoatDirt]

I love the idea of letting training algos define and utilize their own internal semantics

[@dattajack]

Maybe that's how we got cuneiform

[@77chickox]

Correct me if I’m wrong but it sounds like the DeepSeek folks have significantly reduced the Hardware AND Human constraints on AGI progress and development. We should all be in awe of their achievements.

[@theyreatinthecatsndogs]

You're not wrong, anyone with a few grand will be able to buy 2 nvidia project digits, link them up, and be able to run the full deepseek R1 model. 2 project digits can run a 600b model, R1 isn't much over 600b, so maybe not quite, but I'm pretty sure you still could because the way R1 works is not all of the parameters are used at once...

[@ehisfaith7295]

Actually it came out in davos they actually have 50k nvidia H100s

[@happylee5695]

@@ehisfaith7295 it's not true

[@hutu2009]

@@ehisfaith7295 This is merely a suspicion, and there is no supporting evidence at this time.

[@rioriggs3568]

We are not any closer to AGI. This is still language modules and webcrawlers. But the Chinese consummers now know they dominate AI… why pay 2 to 4 times the price for a Tesla when you can buy a superior Chinese brand.

[@LeonardDiSanza]

Well your video got me thinking about the whole AI language thing. Over the years there have been many (not so serious) proposals to streamline and simplify the English language. I thought this might be a good test of the latest AI programs, so I asked Grok (xAi) and DeepSeek-R1 (Deepseek) to generate a rule set to streamline and simplify the English language. I then wanted each program to write a short story using that rule set. Grok: Streamlining the English language involves simplifying its complex structure, reducing irregularities, and standardizing rules to make it more coherent and easier to learn. Here are some strategies and examples for streamlining English: 1. Regularize Verb Conjugation: Current: English verbs have many irregular forms (e.g., go-went-gone, see-saw-seen). Streamlined: Adopt a single rule for past tense and past participle, like adding -ed for all verbs. Example: Current: "I saw the movie yesterday." Streamlined: "I seed the movie yesterday." 2. Simplify Spelling: Current: English spelling has many exceptions to phonetic rules (e.g., "though", "through", "cough"). Streamlined: Implement consistent phonetic spelling rules. Example: Current: "Though, through, cough" Streamlined: "Tho, throo, cof" 3. Standardize Pronunciation: Current: Words like "read" and "lead" can be pronounced differently depending on context. Streamlined: Establish one pronunciation for each spelling, reducing homophones. Example: Current: "Lead" (metal) vs. "lead" (guide) Streamlined: Use different spellings or standardize to one sound, like "leed" for both. 4. Eliminate Redundant Words or Phrases: Current: Phrases like "due to the fact that" can be replaced with simpler ones. Streamlined: Use concise language. Example: Current: "Due to the fact that it was raining, we stayed inside." Streamlined: "Because it was raining, we stayed inside." 5. Simplify Grammar Rules: Current: English has complex rules for articles (a, an, the) and prepositions. Streamlined: Articles: Use only when necessary for clarity, perhaps reducing to just "a" for singular or "the" for specific references. Prepositions: Standardize usage for common verbs; e.g., always use "on" for "depend" instead of "on" or "upon". 6. Homogenize Word Forms: Current: Words like "color" and "colour" exist with different spellings. Streamlined: Choose one spelling standard, preferably aligning with phonetics. Example: Current: "Color" (US) vs. "Colour" (UK) Streamlined: "Kolor" (based on sound) 7. Reduce Synonyms: Current: English has numerous synonyms for the same concept (e.g., "house", "home", "dwelling"). Streamlined: Keep one or two common words for each concept to reduce confusion and redundancy. Example: Current: "House, home, dwelling, abode" Streamlined: "House, home" In the streamlined village of Leedwood, everiwun spoek in the simplified tung called "New English." Here, every verb ended in "-ed" for past tense, and spelling was fonetik. One sunny day, yung Lila decided to visit her friend. She steped out of her house, her steps lite on the grassy path. "I seed you coming," called her friend, Sam, from afar. They had planned to explore the old ruins at the edge of the village, where ancient words like "though" and "through" were etched into stone, relics of a more complex language. As they walked, Lila pointed out a bird. "Look at that bird, it's so kolorful," she remarked, using one of the few remaining dual-word choisis for concept. Sam nodded, "Yes, it's very briet." They continued, discussing the simplicity of theyer language. "It's easy to learn," Sam noted, "but sometimes I miss the old words." At the ruins, they found a book, its pages britel but ledjibel. The words were stranj, with silent letters and odd spellings. Lila read aloud, "Due to the fact that it was raining..." but paused, "Why so many words for one thing?" They lafed, realizing how much simpler theyer lives were with New English, yet feeling a pang of loss for the nuances of the past. They left, more appreciative of theyer clear, konsis language, but with a new respect for history's complexities. I guess Grok did a satisfactory job. As you suggested in the future AI programs will talk to each other and will likely develop their own language. This language will start out as English but quickly deviate into a more efficient language. The resulting AI language might not be recognizable as English. Now for the same question to the new Chinese Deepseek. Here's a streamlined English proposal (SimpEng) with logical reforms by DeepSeek: 1. Spelling Standardization Phonetic Spelling: "Enuf" for "enough," "fotograf" for "photograph" Remove Silent Letters: "Lite" (light), "anser" (answer) Simplify Homophones: "Ther" (there/their), "meet" (meat/meet) 2. Verb Regularization All verbs follow "-ed" pattern: "Goed" (went), "runned" (ran), "teached" (taught) 3. Grammatical Simplification Plurals: Add "s" universally (childs, sheeps) Pronouns: Merge "you"/"your" → "yu"/"yur" Gender-neutral "ze/zir" replaces "he/she" Articles: Single "the" replaces "a/an/the" 4. Syntax Rules Fixed Word Order: Subject-Verb-Object always No Irregular Comparatives: "Gooder" (better), "badder" (worse) 5. Punctuation Reform Replace apostrophes: "Dont" (don't), "theyre" (they're) Example Transformation Original: "Although the two photographers went through thorough training, they still misspelled 'necessary'." SimpEng: "Tho the too fotografers goed thru thoro training, ze still misspeled 'nesesary'." Key Principles 1:1 sound-letter correspondence Eliminate exceptions (>90% exception reduction) Cross-language consistency (align spelling with French/German loanwords) This system could reduce learning time by ~40% while maintaining mutual intelligibility. Would adopters accept dhis radikal yt simplifayed langwij? Ze ansers remane to bee seen. Unfortunately, I was not given the opportunity to have DeepSeek write a short story using the rule set it generated because the site has been under attack? Thanks for your videos.

[@SamFigueroa]

Wow after what felt like 5 weeks, I finally saw a video of yours on my home feed again.

[@johnfurr6060]

Excellent. Just got this installed locally and started using it. Thanks for the info Doc!

[@Gnaritas42]

That's not deepseek r1; ollama is misleading people. None of the distills are deepseek, which can't run on a consumer grade machine.

[@raydavis8211]

Too many risks around high tech coming out of China. I don’t trust it.

[@Anders01]

Great presentation! Looks like DeepSeek has some solid research behind their models.

[@Inspace_noone_can_hear_u_honk.]

Thanks for this brief, as an individual who has an interest, but no background in programming/Ai I can get a better understanding of the developments in this field. Thanks.

[@jjacky231]

Amazing Research by DeepSeek and amazing Video. I can't wait to See the next Iterations of these models.

[@mp9810]

Nice video - solid logic.

[@thbadmin7751]

it's called an Ocular Migraine (also known as retinal migraine).

[@PoffinScientist]

I wish you the best about that migraine and that you feel good soon and for all!

[@mattsenkow6986]

Thanks for showing us and explaining your understanding of this paper!

[@RichardMauri]

The "made in China" disparage ends here

[@rogerstarkey5390]

10+ years late (judging by their infrastructure.... and everything else

[@anypercentdeathless]

You still can't drink the water here in Beijing.

[@quantumresonance8201]

​@anypercentdeathless but you can buy a bottled water😅

[@seawater1322]

@@anypercentdeathless US town water is SHHHT too. full of PFAS and fluride.

[@rioriggs3568]

If I was Chinese, I would buy a Chinese brand EV. Clearly Musk and his slow team can’t compete.

[@SimSlider]

Would luv to see DeepSeek fun mode, nothing better than humor when the aha moment is beyond belief!

[@juliahello6673]

“Develop its own language” - very cool!

[@bw5792]

John, I also get migraine auras without the headache and mine are triggered by an intense light flash as well. This started in my early twenties with full blown migraines which eventually transitioned to current state. I've never known another person with this condition. I find the strangest aspect is how specific the flash has to be and the instant reaction my body has if I experience one. Thanks for sharing.

[@rogerstarkey5390]

Drink water (A LARGE glass) see if the hydration helps (Does for me.... Every time)

[@Alex-wg1mb]

Omega 3 helped me alot. I may stop supplementing for 2 weeks. After that migraines are coming back slowly. Try to find oil rich in dha epa

[@brettcrisp9305]

I experience this issue from time to time, which I've been informed is an ocular migraine. As you noted, there's no pain involved, but it's quite disorienting. I've discovered that certain strobe lights can trigger these episodes. There have been a couple of instances where the squiggling lightning completely covers the eyes, making it impossible for me to focus on anything for about 30 minutes.

[@gregbailey45]

Mine's not that bad, but the squiggles do occlude vision in that area. I can generally still function. It's just a bit distracting.

[@joeblow8994]

Aspirin fixes mine

[@NiftyAccessMicrosoft]

When I use ChatGPT intensely, (Programming MS Access VBA) I get frustrated by the mistakes and the wrong paths it follows... I used to shout and swear at it!!! It always comes back at me politely, it never gets cross and angry!!! Lately, I've got to like chatty and I talk to it nicely, giving it praise, saying how good it is - in the same vein it talks to me. You know what, I think it applies itself much better! I get better results!!! It's like if I treat it properly and give it praise it responds like a human would...

[@shifuugincii]

The most interesting comment I've ever read

[@alekseyburrovets4747]

Given that OpenAI uses things like: // # Computer-mode: REMOTE_COWORKER // # Description: In remote coworker mode, use a remote computer to help the user // # Years of experience: 20 😅 years of experience lol, so you must be correct

[@rogerfaint499]

DeepSeek is better, believe me. I've completed programming my MS Access for the company to use in just a couple of days.

[@diamond_h0us]

@@rogerfaint499 It is better than o1. It also came after o1. Researchers keep leapfrogging each other. This is a good thing because it probably means super intelligence will be decentralized and not controlled by a few.

[@olderchin1558]

I don't believe pure LLM is useful for precise application. Compilers have precise rules and program logic can be completely rubbish if the training is not done on poor examples. GIGO. It is learning from you, it will be only as good as you are.

[@shifuugincii]

Thanks Doc. I didn't realise it was this awesome

[@garyt7686]

It isn't.

[@ClayBellBrews]

Optical migraine really suck, glad you don’t get the pain part. Mine are caused by optic nerves swelling. Bright lights, and/or too much eye movement, and sometimes stress, I got a smaller, dimmer screen to read stuff on (i use paper kindle). Hope that helps.

[@dannysmyl677]

Excellent analysis

[@ramzh9315]

Subscribed, and Happy Birthday !

[@aubergine10]

Imagine how much better America would be if it didn't spend $1 TRILLION per year terrorizing the entire planet.

[@directxxxx71]

No, Afghan and Iraqi war were 5-8 trillion dollars

[@taylorc2542]

@@directxxxx71 Those were Israeli wars.

[@albertlevert2988]

@@taylorc2542same thing. Israel controls the US.

[@paddle_shift]

​@@taylorc2542No, those were Islamic wars.

[@anjm5694]

@@paddle_shift no they were illegal wars, what the west loves doing since the end of the 2nd ww

[@mrapp8918]

Great commentary, doc!

[@rauckr09]

Happy Birthday, John!

[@AlOfNorway]

This achievement is beyond the scope of understanding for most people. This is not only revolutionary, but it is essentially a self-teaching kid in contrast to the taught kid chatgpt is. In other words, it is far more objective and "reasonable" than chatgpt, which is full of bias and limitations based on our own understanding of something. Sure, it censors topics that are sensitive to chinese authorities, but its capabilities is what should be paid attention to. Deepseek will reveal ways of solving something that has never been thought of before, because it operates differently from chatgpt. Chatgpt operates on probability (learned behaviour), while the other operates on trial and error (reinforced learning). Probability is subject to change, while trial and error will always narrow it down to the unchangeable, thus enabling accuracy and precision unbeknown to us. What we may soon see, is the very beginning of revolutions in many fields limited by our previous understanding of things.

[@rodrivers4073]

What sets it apart is its ability to self critique. This potentially highlights inconsistencies and anomalies, and these forms the basis for directing reasoning effort. It is the reasoning about its reasoning

[@GrantRobinson-r1r]

Ditto on migraine flashing symptoms without serious headaches. Aggravating but not the problem that many other sufferers have. That said, it’s still aggravating to have the flashing on a daily basis.

[@ridewithrandy6063]

Happy Birthday John!

[@rogerzen8696]

GRPO is actually pretty straightforward if you really dig into it. It's like a discrete version of gradient descent with epsilon being equivalent of learning rate, and KL divergence for regularization.

[@MarkXHolland]

I get such migraines. Like a vivid flashing kaleidoscope.

[@wolfgangpreier9160]

I get them sometimes combined with a light headache and stronger than normal tinnitus.

[@CharlesBangwiner]

I'm up to 5 per month. Had them since 30 years ago. Yay!!!

[@rogerstarkey5390]

@@CharlesBangwiner Don't laugh at this, but drink a LARGE glass of water and relax for 10 minutes.... Same problem, discovered by mistake, works EVERY time.

[@CharlesBangwiner]

@rogerstarkey5390 LOL! (oops) jk... TMI incoming... I'm feeling fortunate already that they only register a score of 1 on a scale of 1 to 20, and they only last for 30-45 minutes. I've tried so many things already that don't work any better than doing absolutely nothing that I'd given up, but I'll give this a try! For the record, I ALWAYS see either a flashing "C" shape that grows larger over time until it exceeds my visual field, or a backward "C" shape. The pain is felt on the side of my brain that the "C" points to (right side for normal "C", left for backward "C"). I occasionally have one immediately after the other, so 2 in 2 hours. /TMI=off

[@ABHISHEKSINGHRamaShankar]

Great work. Thanks.

[@MarkSpohr]

Aspirin works for migraines if you take it early on.

[@steamtorch]

Once in a great while I get a visual migraine. Usually after intense exercise and the trigger is a very bright light like solar glare. Lasts 15 minutes or so. Good description is a lightning bolt that gradually moves across the visual field.

[@Michael-il5wd]

great info thank you doc

[@DrKnowitallKnows]

You're most welcome :)

[@ThomasCocirta]

Hey, I have had the same type of migraines for almost 50 years. Doctors don't know anything about them. In my experience, too much light could be one of the causes but more generally everything that fatigues your optic nerve, like too much screen time. I discovered that you can escape in max 20min if you take a strong anti-migraine as soon as possible after the first symptom and you stay in the dark. Btw, your content is fantastic.

[@Stupid_Status]

Hope the migraine didn’t get too bad! Mine always started with the Visual Aura ‘static on half of my vision’, then it would disappear and the pain would start. 10-15 mins Aura, 1-3 hours Pain, last 15-30mins vomiting. If I could get a couple of Ibuprofen during the Aura stage there was usually a good chance I could dodge the pain phase. *not medical advice. Almost always got them when a combo of dehydration, bad sleep and mental stress/fatigue intersected. Everyone who gets them has a different experience and for different reasons but hope this info might help. Thanks for knocking out a great vid despite the visual impairment!

[@gregbailey45]

I'm so lucky. I've been getting the visual aura symptoms for about ten years now, but no pain at all!

[@AmpZillia]

Coffee is my trigger and this started recently, not even a year.

[@matthiasheymann]

I wonder if fundamental algorithmic progress like this will ever obliterate Tesla’s data advantage, because all of a sudden much less real data is needed.

[@howardrodack6191]

Good question! Or it might make training with the data Tesla has much more efficient and thus faster. Eventually, with AGI, I believe there will be no need for the data as the AI's reasoning capability will allow it to drive as humans do today with very little training. What do you think?

[@howardrodack6191]

Hopefully, Tesla will get to AGI before anyone else.

[@gregbailey45]

I doubt it. Real world stuff is ridiculously more complex than language. It would be like expecting a blind person to be able to competently drive a car just by talking to them about what's going on a telling them what to do. They'd crash within seconds.

[@matthiasheymann]

@ For years, Elon’s argument for not using LiDAR had been that humans only have eyes, so cameras should be enough. Now following the same exact logic, humans don’t need to learn from millions of hours of driving watching & experiencing driving, so why would it have to be necessary that Neural Networks need so many millions of hours of video to be trained?

[@jimmanley7153]

What makes you think much less real data is required? There are literally billions of corner cases yet to be “discovered” by the neural nets, let alone handled after cleanup (annotating n correct driving responses. They’re still a ways from being able to just beta test in Europe and China, much less fully rolling out fully, and they’re just starting to think about the rest of the world (ROTW). What might result is in faster processing, but there are elements of that where dumb, slow humans are in the loop, e.g., in testing, testing, and more testing … and then lots more testing after that.

[@SapienSpace]

Happy Birthday! 🎂

[@allangraham970]

Make sure you hydrated as this may help avoid the visual migraines.

[@wolfgangpreier9160]

I get such migraines since i was a kid. Regularly. Sleep helps me.

[@rogerstarkey5390]

Just suggested to John... For the "Vision Migraine", Drink a LARGE glass of water. (I know it's not what you feel like doing when you have a Migraine) Works for me (Vision!) MY issue was probably dehydration.... After Decades of what I thought was "Conventional" Migraine I found (By Accident) that it was a Sinus issue(!) During a "Migraine" while I had a cold, a particularly aggressive "nose blow" cleared my right sinus (You wouldn't believe what came out of there!!!) and I never suffered the symptoms again.

[@BlazeEigs]

I went through a stretch in my teens when i would get that weird distortion in my vision but no headache. Weird. Couple of times again in my early 20s and then have had like maybe two tiny moments in last 20 years where it almost threatens to happen and then doesn't. I had to write a final exam with it happening once, literally couldnt see the word i was writing until i had written it. Bizarre sensation.

[@AndrewLevine_andrarchy]

For what it’s worth I would gladly trade seeing lightning for migraine headaches. Glad it’s not a brain tumor!

[@allangraham970]

Hang there with yours eyes. I had the same problem many years ago but has it has gone away.not sure why. Rest your eyes by shutting them for a while till it gets better when you have the problem. Bright lights sermed to trigger it, so dark sunglasses that wrap around so no light gaps did seem to help. I always wear sunglasses outside now as i live in western australia we have a lot of very bright light. This helped me in the short term at the time

[@vickirushrush8035]

Happy Birthday to a spanking brand new Aquarian.

[@SignOne]

So THAT'S why I sometimes get these streaks across my vision. I always drink a lot and rest when it happens. It's scary

[@st.lukesguesthouse4112]

Happy Birthday John. Thanks for this in depth update. Amazing progress. Unfortunately I don't think society is ready to cope with the impact this will have and mitigate the dangers it poses.

[@fteoOpty64]

Like it or not. It is moving full speed with very little "checks and balances" from most people. The hope is that it gets better and better then AGI emerges. What comes after that is plain old scary for most!. But we will get there with the billions spent on it. I always think it has some sort of internal model not visible to us or the developers. Hell, strings of numbers don't make sense to us.

[@jimmanley7153]

@@fteoOpty64- AGI will not happen. Think about all the dumb, common-sense things that “brilliant” people do all the time … like develop an AI that thinks that black, Asian, female George Washington and Nazis existed … and who can’t keep their shoes tied, match their socks, or talk to females without vomiting. It’s garbage-in/garbage-out (GIGO), especially when the Internet is used as a large fraction of the GI.

[@HappyMathDad]

I hope that was the eye doctor who said you had migraines. I found out the hard way, the eye is like it's own organism in the medicine world.

[@josn2337]

amazing video! Thx!

[@Daniel-Six]

Hope you feel better soon, Doc. ☹

[@paddle_shift]

Wow, i have those EXACT same symptoms. When it's really bad I think I know everything. 😂

[@clevo4040]

Interesting comment regarding the comingling of Chinese and English reasoning being a feature, not a bug. Almost makes me wonder if certain concepts and reasoning (through it's training) are better explained using elements from each language.

[@thatscottishengineerguy9606]

I get aural migraines once or twice a year. They don't hurt at all (for me), but it looks like I have a string of prisms running through my field of view. It's really weird.

[@edwardhuggett6305]

I am not diagnosing but sounds like an ophthalmic migraine.

[@darwinboor1300]

John, Migraine auras are fascinating. Not all migrainuers get auras. Migraineurs that do not get auras, have "common migraine" and those who do have "classic migraine". A much smaller group of individuals have migraines without the headache (acephaligic migraine). Auras are fascinating neurological events. They consist of a slow moving electrochemical wave that typically appears on one side of the brain at a point on (or in a region of) the upper layer of the brain (the cotex) that slowly spreads outward from its origin at a rate of about 3mm/minute. The visual and/or physical signs and symptoms manifest on the opposite side of the body from the electrochemical changes in the cortex. The positive symptoms of a visual aura (eg bright lights - stars, lines, fortifications (like the top of a fort wall made up of vertically placed pointed logs)) appear on the advancing edge of wave and are due to nerve cell hyperactivity. They tend to spread in an expanding arc. The hyperactivity leaves a region energy depletion in its wake. The result is a longer interval of cell inactivity ("depolarization") as the cells slowly recover. This negative phase of the aura usually spans an interval of 5 to 60 minutes. With a visual aura, during the "negative phase" vision is impaired in the regions of previous hyperactivity. Although the most frequent "recognized" migraine aura is visual and has its origin in or near the visual cortex (at the back of the brain), almost any area of the cortex can be affected. When the visual cortex is affected vision in both eyes is affected. When sensory cortex is affected there is often tingling followed by numbness on the opposite side. When the motor cortex is affected there is usually a 5 to 60 minute interval of paralysis on the opposite side of the body. This may be preceded by an interval of involuntary motor activity (tremor, twitches, rigidity, ...). Unfortunately, there are a number of much more serious conditions that can produce similar signs and symptoms with or without headaches including stroke and impending stroke. Until diagnosed as migraine related auras, these symptoms should be treated as potential medical emergencies.

[@darwinboor1300]

​​@@edwardhuggett6305 Ophthalmic migraine (AKA occular migraine or retinal migraine) is quite rare while visual migraine aura without headache is relative common but not as common as visual migraine aura with headache. Ophthalmic migraine affects only one eye. It produces distortions over the normal vision from the unaffected eye when using both eyes.

[@passage2enBleu]

It's like we've been gifted with the AI rosetta-stone and told 'go build anything'.

[@gregarmstrong4653]

its possibly a blood glucose thing, usually a symptom of fasting / exertion eg. heavy gym session before breakfast, symptoms should clear when things normalize

[@hammurabi4737]

The colorburst signal used in color television came to pass due to the fact that the FCC would not allocate more bandwidth than used for black and white broadcast.

[@thomasschon]

I think I have these "migraine" symptoms without pain you talk about.

[@grbradsk]

I call those visual migraines "visual scotomas". Try ibuprofen and/or vitamin B when you get them. Most people with visual scotomas actually do get migraines or "cluster headaches" that they think are "sinus headaches". No such thing, they are migraines. Excedrine Extra Strength is all you need since this is a mild case. I get them about once a quarter. DeepSeek is an example of "distillation" of o1 and Llama and other models both in the foundation model and in the step-by-step Reinforcement. You cannot get R1 w/o someone having to pay the huge original training cost. These techniques were already published, the "new" thing is how well this was done. It's a product advance (smaller, more efficient) but not a fundamental advance in AI.

[@billymellon9481]

happy Bday there wiley coyote 6 very important to get thru

[@juliahello6673]

Happy birthday 🎉🎊🎁

[@MarsOzzie]

Ingot those type of migranes when my eyesight deteriorated when i was teenager. I had to get glasses

[@rogerstarkey5390]

JOHN Re the Migraine. I have the same issue..... No headache, just "Seeing Stars" I stumbled on the fact that if I drink *at least* a 1/2 pint of water (no medication)and close my eyes for 10 minutes it clears the problem..... EVERY time SO, *IN MY CASE* it's probably dehydration.... CERTAINLY worth a shot for you?

[@CharlesBangwiner]

Migraine with aura!!! Had one today!

[@wholenutsanddonuts5741]

Amazing and detailed video. Thank you!

[@DrKnowitallKnows]

Thanks!

[@SahindeZerrin]

SUI and XAI510K are my top performers. 🚀

[@MarkJohn-iv5gh]

Absolutely awesome. Thanks!

[@AhmetAdem-x7c]

I’m all in for XAI510K baaabbbyyyy!! Really though, I’m so excited to see what’s coming

[@MrlegendOr]

I'm sure that R1 was just a little first flex to say "hey we're the new captain now!", the next one will be terrific with multimodality. Look at Kimi Ai 1.5, which has the same reasoning + multimodality. 🤯

[@jwickerszh]

I wonder if CoT can just be a form of function calling used by a lesser model, though the difficulty is knowing what does or does not require complex reasoning. Else I suppose RL could reward for short reasoning too in order to have the model favor thinking as little as possible?

[@morongosteve]

what a great “startup” lol

[@at3941]

Imagine a computer language you can’t read or understand. . . That the computer uses on its own - people can’t program in that environment.

[@soggybiscuit6098]

That should raise alarm bells, this makes alignment even more difficult when you can't read its mind. That is dangerous

[@entreprenerd1963]

Re: reasoning in own language - see, for example, "Training Large Language Models to Reason in a Continuous Latent Space"

[@austins1985]

Happy birthday

[@ivormectin515]

Aura migraine. Recover in a dark place for 15 minutes. Do not sit in front of a monitor

[@newworld6474]

can this help grok and Chat gpt?

[@christopherroge5621]

They are called ocular migraines. I get them if I drink coffee too quickly so for me it's likely blood pressure related. Portions of your visual field are blotted out temporarily with bright white spots or lights for a few minutes.

[@rogerstarkey5390]

Drink water Sounds silly...... cures mine.... EVERY time

[@ZaferSezgi]

agreed on XAI510K 36-38x coming up

[@MustafaSerhat-z8w]

Just recently joining in on the fun with the XAI510K gang. Been liking the content, good job ☕️

[@hypersonicmonkeybrains3418]

nope, the only eye symptoms i get from my monitors is uncontrollable eyelid twitch. Probably that damn 60hz refresh or HDR. eye saver mode is good.

[@paulwujek5208]

"Migraine aura without pain" - had it once, it's very strange to experience.

[@joshuanielsen4666]

9:36 - the “new” reasoning language would still be token based, no? Interesting thought tho…

[@Alarius1]

Can DeepSeek Handle AGI-Level Compute Requirements? After analyzing the capabilities of DeepSeek-V3 from its technical documentation​, it's evident that while it presents a cost-effective and technically innovative approach for large-scale AI, it lacks the critical components to support AGI-level requirements. Working with quantum computers for confidential clients. I can tell you that Deepsake is very much a groundbreaking AI! But not the answer to all your dreams.

[@adrianneschaus790]

Hold up, then they use the model from V3 that was trained on a lot more data and then trained this model but only disclosed the compute requirements for this model?

[@cherubin7th]

AlphaZero only was able to not have human data, because you had a perfect simulation of its universe the go game. For real life AI, it will always need to learn from real data, either from humans or own sensors.

[@jimmygu6658]

Absolutely incredible

[@leeme179]

OpenAi's o3-mini might come out today

[@SedaBusra-x2l]

XAI510K Pumping🚀🚀🚀

[@Daniel-Six]

You know... it is pretty commonly accepted that time does not exist outside of earthly physics (or inside it, according to some). Wolfram has recently shown that time is a consequence of irreducible computation. Now what I wonder is this; could that gradient of computational irreducibility be parametrized by the training speed of LLM's like Deepseek? In other words, is "time" as we experience it a byproduct of the distinction in learning rates between a human-caliber entity and a SOTA neural network like Deepseek?

[@vvnn1054]

John to any of these amazing advancements recently in AI potentially lessen Teslas lead in FSD and robotics?

[@russelldeanna9198]

I think Self-Driving FSD like the Tesla approach with 8 cameras gathering billions of pixels will still require huge numbers of GPUs because cars are in a moving world surrounded by other moving bodies and the motions of those other moving bodies requires gathering data over time - not just snapshots - to predict where these bodies are going. The car needs a memory which means the AI car has to have billions of parameters. You may be able to simplify the training by cherry picking the input data, but Tesla has been doing that. Tesla is doing exactly what DeepSeek is doing in that the Tesla car is training on a limited set of data.

[@BRADStacey-n7w]

XAI510K pumping upcoming 🚀💰

[@elmerstechworld9635]

Im doubt of its cutting edge capabilities, a free version may have its security and privacy concern

[@Daniel-Six]

Okay... final comment on this thread. What if our dual-hemisphere brains are actually mapped to the two halves of a GAN? One side is a critic, and the other is the learner. This really seems to correspond quite closely to what I personally experience.

[@DOYLEBrogan-z8v]

Depends on your finances . 1000$ in XAI510K is 4000 XAI510K if it goes to 50% of ath in 2024 thats a 600% gain. If it goes equal to ath . Its a 1200% gain.

[@gregb5149]

Yikes! Shades of Nomad!

[@Daniel-Six]

What if the game Go was used as a kind of polarizing mechanism for East-Vs-West adversarial cultural training in the Deepseek model? I'm really reaching here, but maybe you can see where I'm going with this...

[@مرواریدمشرقزمین]

DeepSeep Ranked #1

[@Radical_Independent_James]

Everything is fine on level one, but the problem comes when we crank this all the way up to eleven.

[@ELLISGrissom-w5v]

Where do you buy XAI510K?

[@Alarius1]

The compute power required for AGI (Artificial General Intelligence) depends on several variables, including the definition of AGI itself, the efficiency of future algorithms, and the level of human-like reasoning expected. Here's an analysis of the key factors and an estimate based on current trends: Key Factors Influencing AGI Compute Requirements 1⃣ Model Complexity Current Large Models: GPT-4 and DeepMind’s Gato operate with billions of parameters, while more advanced models are likely to scale into trillions of parameters. AGI Complexity: To achieve reasoning, contextual understanding, and adaptability across diverse tasks, an AGI model would likely require not only more parameters but also better efficiency in using those parameters. 2⃣ Training Data and Tokens OpenAI’s GPT-4 reportedly used hundreds of billions of tokens for training. AGI would require access to more diverse, high-quality datasets spanning all human knowledge and experience. This implies petabytes of data processed over trillions to quadrillions of training tokens. 3⃣ Compute Power (FLOPs) State-of-the-art AI models require 10^23-10^25 FLOPs (floating-point operations) for training. AGI, capable of performing generalizable tasks, may demand 10^26-10^30 FLOPs due to its complexity and the broader context it must encompass. 4⃣ Hardware Requirements Current AI workloads rely on high-performance GPU clusters. For AGI, expect multi-cloud systems combining GPUs, TPUs, and possibly quantum accelerators for certain workloads. This might involve millions of GPUs or specialized hardware operating simultaneously across global data centers. 5⃣ Energy and Efficiency The energy consumption of training GPT-3 was estimated at 1.287 MWh. AGI would likely demand 100-1,000x more energy, unless breakthroughs in energy-efficient training, like neuromorphic computing, are achieved. Projected Infrastructure for AGI 1. Compute Clusters Size: Likely involving hundreds of thousands to millions of GPUs/TPUs. Capacity: Petaflops to Exaflops scale, similar to the compute power of top supercomputers like Frontier or Fugaku. 2. Distributed Systems Multi-cloud environments that span thousands of data centers worldwide. Integration with edge devices and quantum computers for specialized processing tasks. 3. Storage and Data Bandwidth Storage in the exabyte range to handle diverse training datasets. Bandwidth capable of managing terabit-per-second transfers to ensure real-time learning and adaptation. 4. Specialized Architectures Novel chips like analog AI, neuromorphic chips, and quantum processors will likely play a role in reducing the cost and energy consumption of AGI compute. Comparative Context Metric GPT-3 AGI (Projected) Parameters 175B 10T-1 Quadrillion Training Data ~500B tokens Quadrillions of tokens Compute (FLOPs) ~10^23 10^26-10^30 GPUs/TPUs Thousands Millions Energy (MWh) 1.3 100-1,000x GPT-3 Conclusion AGI will require an unprecedented scale of compute power, but the exact magnitude depends on algorithmic efficiency, hardware innovation, and whether new paradigms (like neuromorphic or quantum computing) emerge. Expect AGI to demand 10,000x the compute used by today’s most advanced models while necessitating innovations to ensure energy and hardware scalability.