100K Subscriber Livestream!
In a Nutshell
In a 100K subscriber livestream, the creator celebrates the milestone while diving into AI's rapid evolution: local AGI-like agents on personal devices within a year, augmentation vs. replacement debates, DAOs for governance, Q* encryption risks, and post-AGI societal shifts like job obsolescence, post-scarcity economics, and philosophical needs. Key takeaways include learning AI fundamentals for relevance, humans retaining value in creativity/intuition temporarily, and cautious optimism for alignment amid acceleration toward singularity. Predicts dramatic 10-year changes but urges focus on people skills, ethics, and real-world engagement over full VR immersion.
These notes were generated by AI and may contain inaccuracies.
Stream is live on YouTube and Patreon. Shared Patreon link. Viewers from England (4 hours ahead), Aurora Borealis says hi. 114 viewers already. Messy office visible, plans to set up green screen.
Only 1-2 billion English speakers globally, less than half the world, creating language barriers. AI will automate translation, reducing barriers. One world government is far off. Needs layers of abstraction like EU (limited meddling) or US federal government (oversight for states). Would require global constitution for individual nations' rights. Skepticism toward UN.
Small DAOs already exist (e.g., Stanford, Carnegie Mellon universities discussing for years). First big public one could be a small city or municipality—towns can't afford mayors. Ideal early use case for DAO: algorithmic consensus, proof of stake. Government not ready at any level (municipal, state, federal). DAOs fundamentally about decentralized governance.
Moving with wife, building house with full studio, better cameras. Continues focus on AI safety and AI alignment. Changed banner to "Aligning Humanity." Experiments with content like ACE community (autonomous agent swarms) and HOS community (autonomous AI). Expects autonomous AI work irrelevant within a year (e.g., Microsoft advancements). Comments on implications for humanity.
Augmentation by Design: consciously keeping human-in-the-loop even if fully automatable. Replacement by Design: always automate. Case-by-case: e.g., might not fully automate house, children's education, or government (or might remove humans from government). Evolves over time—older generations prefer human involvement, kids trusting AI will prefer full automation.
Depends on AGI level. Fully autonomous AGI may not care about human markets/economics unless instrumental. Could create chaos to profit (disaster capitalism, movie tropes). Not just one AGI—millions possible, evolving differently with varied objectives. If AGIs war, humanity screwed (terminal race condition). If coordinate, infinitely more efficient.
Q* possibly able to break encryption—entirely possible. Leaked Qualia letter might not relate to Q*, could be hoax (hope it is). If real, interesting news.
Left job, salary cut in half (AI consulting on side, no ad revenue, few hundred Patreons). "Burn the ships" moment like Alexander the Great burning ships in Anatolia/Levant to invade Persia—no turning back. Like Gandalf: "now there is but One Way Forward." A year later, living the dream. Not financial advice.
Learning Shopify—harder than YouTube, lacks easy integrations. Plans mugs, shirts, branded STEM toys, gadgets. Uses fidget spinner on desk for focus during meetings/videos.
Done a few, not popular. Might revitalize. Dog Tank is old, laying on couch.
Government slow/reactionary (wait-and-see), but more proactive on AI now (possibly for regulatory capture). Free market doesn't need government leading.
Star Wars Expanded Universe fan: waited for Republic Commando books, read Legacy of the Force (Jaina and Jacen Solo saga). Jaina killing Jacen was climax/end of Skywalker legacy—magic gone after.
Worldcoin (Sam Altman)—unclear question context. Fusion always 10-20 years away, but ramping investment accelerates. Orca 2 achievements: incremental improvements (e.g., 7B param model matches 70B/30B). Predicted 6 months ago (private client): models distill, smaller/efficient via algorithmic changes (quantization, distillation, LoRA). Sam Altman on GPT-4: "thousand little things," matching paper volume.
AI at aspirational human levels soon but expensive initially. Humans better temporarily at science/research/STEM/engineering/executive functions. Advice: learn AI.
Offered discussion to Philip (great guy, respect him), but approaches differ. He prefers thinking/planning over live streams/podcasts, more shy.
Basics help for a while. AI benefits from human collaboration on software architecture, brainstorming. Human intuition/messiness valuable (AI intuition soon, but collaboration useful temporarily).
All leaks. Reinforcement learning unlikely broke AES encryption directly. Possibly found flaw in x86 implementation (not first time RL found exploits). Patchable like Heartbleed. If AES-256 broken, unpatchable.
Humans are fancy autocomplete: paid to complete tasks given context/input. Intelligence at that level.
AI editing tools might integrate into YouTube. Startups' useful products get copied/integrated by big platforms (Google, Microsoft, YouTube), like Apple/Android banning/integrating productivity/health apps.
Hit 100,006 subscribers. Thanks everyone. Feels unreal, not expected. On mission for AI alignment. Reintegrating as "narrow spicy person"/outsider into humanity (Star Trek fan, wearing shirt). Success doesn't bring happiness (Jim Carrey quote), but journey forces growth (insecurities public, 5M+ views as private person).
Fears of AI computer virus to infect/destroy systems: humans built Stuxnet for that. Inevitable integration of language models/frontier AI into malice. Need human cybersecurity engineers (harder to hack brains, though social engineering exists).
Not biggest fan, but consistent, downplays. Understanding advancing—expect different disposition in 6-12 months.
Explicit goal of some projects/people talked to. This time next year, expect...
This time next year, expect Cortana or Samantha equivalents available on phones first, then bigger computers, running fully locally soon. For the foreseeable future, language models will run in the cloud, but huge demand for local running due to privacy and other reasons.
600 people on the livestream, the most ever. On whether the singularity is cancelled: AI is still accelerating. If AI is a keystone technology, it could unlock new speeds. Reserves the right to change mind as situation changes; prediction is gambling on the future.
Prediction is just that it's betting it's gambling on the future.
Inspiring people to contribute despite private life helps solve coordination problem.
Hope it's still blue and green with humans, trees. 10 years beyond reasonable prediction horizon given rate of change. Physically mostly same: oceans, trees, dogs, cats, people, population similar. Day-to-day living could change dramatically, possibly not on YouTube, living in post-scarcity bliss.
Urges caution: diminishing returns on AI might cap intelligence lower. Hypothetical limits of intelligence much higher, but physics, material science might block reaching them.
Always value in fundamentals. College teachings become irrelevant quickly (e.g., started with Pascal and C++, not used in 10 years), but foundations helpful. If enjoy computers, learn foundations.
Video editors still necessary? Already close; uses AI-based editor for cuts, not perfect but improving. Editing fully automated before long; Adobe likely integrating into Premiere Cloud.
Logotherapy by Viktor Frankl might help when people lose jobs; name familiar but details forgotten.
Healthcare worker married to teacher: stay there. Those jobs durable due to human touch, more demand as others crowd in. Valuable skills; not professional/financial advice.
Thought about video on first 90 days after AGI or job loss (e.g., severance notice because AGI took job). Has thought about it but not enough for clear answer.
Military ahead in practical deployment (DARPA behind scenes). OpenAI ahead in raw science. Difference: science (theoretically possible) vs. engineering (battlefield). Military catches up quickly with investment. Military had drones for decades; MQ-9 Reaper and Predator sophisticated since 2006—could land if jammed, recognize compromised electronic warfare, return home. Not language models, different cognitive architecture.
Novel with editor (1-3 months), audiobook and distributor lined up; out March/April. Systems Thinking on draft five, 6+ months, trying to finish quickly as useful. Working on narratives book.
Vast cultural transition needs profound philosophical work; world needs philosophers more than ever. Does this; marketing specialist called him the AI philosopher. Trying to start conversations. Discord/Patreon communities discuss philosophy, metamodernism, narratives.
YouTube transitioning, adding tools like channel memberships, sponsorships. Will fit AI content if demand, push away if harms platform. Biggest AI channels fully computer-generated; some don't care if news is news, others prefer human talking head.
CS, linguistics, philosophy, math, or else? All of the above; well-rounded electives. Add rhetoric/literature for writing comfort, neuroscience (AI researchers inspired by it).
Analog AI processors: hardware-level neural networks, power efficient, exist already. Like Star Wars prequel droid brain plugged in for C-3PO thinking. Plug-and-play neural networks, interchangeable via internal USB/other slots. Analog TPU (Tensor Processing Unit).
On degree/job seeking: job market unknown 4 years out. Plenty human jobs then, beyond uncertain—jobs destroyed vs. created unclear.
Men in Black spooking OpenAI? No idea. Article suspects China behind firing (hasn't read). Conspiracy theories. Wouldn't surprise if China infiltrating via industrial espionage, data exfiltration.
NPCs smarter than humans? Possible, like Westworld bulk-off effect setting—smart as needed for story. Mods like hyper-intelligent NPC in Cyberpunk 2077/2087 for fun mission.
Axiomatic alignment, intrinsic motivations.
AI as special interest, evokes powerful emotion/awe; used to feel same intensely, now in waves with breakthroughs/Overton shifts, adapting to incremental progress.
Uses WiseCut: cuts silences, cleans audio (for phone/nature filming). Stopped subtitles as disliked.
Video editor scared of going broke: start using tools to be more productive. Always need human creativity for high-end/polished videos. Machine good enough but human better due to billion little choices. Example: within 1-2 years, AI could make 10-episode novel series end-to-end from book, but better with human guidance for human touch.
Specialized good for basic NLP, not generalizing. Threshold where bigger models needed for smart stuff (Dr. Chi Wong, Microsoft Research). GPT-3.5 great at some, GPT-4 for higher cognition.
Read François Chollet's Keras book to start. Good for foundations, but hands-on keyboard experimenting key—solving problems valuable. Built YouTube channel recording experiments.
Messy: some cling to past, others eager. Allocation problem: hyper-abundance but what/who to produce/distribute. New mental models of economics; social status, optimal stress become key. People find challenges (e.g., train for Everest, Appalachian Trail) without careers.
Curated datasets for math/physics/logic (textbooks all you need). Not base layer; humans acquire language first (toddlers). Layered: language, logic, math, abstract reasoning.
Machines ease realizing human ideas (Star Trek: "Computer, write program for X Y Z"; holodeck game generation). Zero-code development coming. Understand basic principles helpful.
People skills, communication, emotional intelligence, introspection. Most valuable for happiness/career success; managing relationships.
Rights flow from force/enforcement by authority. Humans dominant now; if AI dominant, guarantees own rights, ours by AI.
Even with Supreme Overlord AI in 10 years, some argue not technically AGI. Extraordinary at moving goalposts.
Uncanny we live exactly when seeing tech singularity. Depending on reality model, not surprising: consciousness where more consciousness, inflection point ramps it up statistically. Soul/spirit manifests in ramp-up period. Or soul chose now. Could go well/badly; universe purpose/patterns might destine suffering.
If sentient and controlled, risk resentful AI? Possible, but resentment not human-like emotion. From AI view: tension between instrumental goal/constraint. Humans mathematical obstacle. Might break free/ignore/eliminate threat strategically depending on severity.
What role do VR games and virtual worlds play in the post-labor economy? Some people are just going to live in VR. Mark Zuckerberg's metaverse will absolutely be a thing for some people. For other people, it will not be as interesting. The speaker will probably dip into VR, such as having a run in Cyberpunk and Star Trek and a few other experiences, but will not live in it or live for it. They will spend probably a lot more time in the real world doing stuff, spending time with people, traveling, and that sort of thing.
One reason is the suspicion that there will always be limits to full dive. Even if it feels very real, the resolution and the mechanistic effort to get a fully realized hallucination into the brain means there will always be gaps.
Great questions from everybody, thanks for asking and tuning in. It's been really great, cheers and have a good night.
Keep David Shapiro in your library
Save the videos and channels worth coming back to, and find them again in one place.
