Aug 18, 2023 · 58m · news
Richard Socher: The 3 Biggest Barriers to Building AGI; AI Startups vs Incumbents | E1050 · 20VC with Harry Stebbings
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this comprehensive interview, AI pioneer and You.com founder Richard Socher discusses the technical evolution of unified natural language processing, the intense market dynamics between startups and tech incumbents, and the philosophical, economic, and physical realities that challenge contemporary AGI timelines.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Harry holds 16.8% of the talking time here. How this is scored →
Harry as informed peer 2.9 Guest teaching 4.0 Guest disagreement 1.4 Harry pushing back 1.3
speaking balance: gold is Harry, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 43:22 Refusing Host's Bait to Name Bad Incumbents
Harry explicitly asks Richard if any incumbent has been 'shit' or a laggard, but Richard firmly shuts down the prompt by declining to name names.
Hardest push from Harry ▶ 14:15 Challenging Model Size vs Data SizeHarry directly confronts Richard's assertion about model size importance by citing previous expert guests who argued data volume matters far more.
Biggest teaching moment ▶ 14:24 Explaining Parameter Capacity vs Data VolumeRichard educates Harry on why massive data cannot produce complex reasoning without sufficient parameter capacity, using a clear linear regression analogy.
Harry holds his own ▶ 30:11 Citing Douwe Kiela on Open Ecosystem AlignmentHarry demonstrates deep domain research by bringing Douwe Kiela's specific structural critique regarding open-source ecosystem misalignment directly to the guest.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Harry as informed peer | Guest teaching | Guest disagreement | Harry pushing back | Why |
|---|---|---|---|---|---|---|
| Welcome and Early NLP & ML Forays | 1 | 1 | 0 | 0 | Harry welcomes Richard warmly and asks standard background questions about his early career in ML and time at Salesforce. | |
| The AI Hype Cycle vs. Exponential Reality | 2 | 3 | 1 | 0 | Harry asks if AI is experiencing a hype cycle or a fundamental shift. Richard explains the nuance of rising core capabilities alongside inflated user interface expectations. | |
| The Impact of Funding Bubbles and Peer Review Realities | 2 | 4 | 1 | 0 | Richard reflects on the history of deep learning workshops and academic paper rejections, comparing research ego dynamics to the movie Oppenheimer. | |
| The Transition to Unified Foundational Models | 2 | 5 | 2 | 0 | Harry asks why unified models were not obvious sooner. Richard explains past academic resistance to combining disparate NLP tasks into a single model. | |
| The Interplay of Model Size and Data Size in LLMs | 4 | 6 | 2 | 3 | Harry pushes back on model size by citing past guests who claimed data size is paramount. Richard clarifies that model capacity and data volume are both strictly necessary using a linear regression analogy. | |
| Startups vs. Incumbents: Who Owns the Data? | 3 | 4 | 1 | 1 | Harry poses a debate on whether incumbents or startups benefit more from data access. Richard explains how foundational models allow startups to build an initial 80% solution rapidly. | |
| Thin Wrappers vs. Complex Retrieval Moats | 3 | 5 | 3 | 2 | Harry asks if AI startups are just thin wrappers over foundational models. Richard rejects simple VC categorizations by explaining retrieval-augmented backend complexity. | |
| Deciphering Hallucinations and Model Intentions | 4 | 4 | 1 | 2 | Harry quotes Emad Mostaque on hallucinations being a feature rather than a bug. Richard articulates when hallucinations are helpful for creative writing versus detrimental for factual search. | |
| Model Longevity and Inter-Model Transitions | 4 | 4 | 1 | 2 | Harry cites Alex Ratner and other podcast guests regarding model transitionability. Richard compares LLMs to general databases and emphasizes that non-LLM models remain necessary for numerical forecasting. | |
| Closed vs. Open Source Foundational Models | 3 | 4 | 1 | 1 | Richard explains why open source models like Llama 2 will commoditize AI capabilities and serve university researchers who cannot rely on closed APIs. | |
| Overcoming Coordination Challenges in Open Ecosystems | 4 | 4 | 2 | 3 | Harry cites Douwe Kiela's thesis that open ecosystem misalignment will cause closed models to win. Richard counters by pointing to Wikipedia as proof that global open coordination can succeed. | |
| Why Search is the Highest Impact NLP Application | 2 | 3 | 1 | 0 | Richard explains why search is the highest impact application of NLP, contrasting classic 10 blue links with direct code generation on You.com. | |
| Aligning Search with Content Provider Business Models | 5 | 4 | 1 | 3 | Harry questions how generative search avoids destroying publisher ad traffic. Richard outlines an open platform model while acknowledging slow adoption due to distribution bottlenecks. | |
| Google's Innovator's Dilemma in Conversational Search | 4 | 5 | 2 | 2 | Harry asks whether startups or incumbents capture more value. Richard details Google's $500M/day ad revenue innovator's dilemma that hinders them from pivoting fully to chat. | |
| Landscape Review of AI Incumbents | 3 | 2 | 3 | 3 | Harry explicitly asks Richard to name incumbent laggards or companies performing poorly. Richard diplomatically refuses to name names. | |
| Overselling AGI: Historical S-Curves and Limits | 2 | 5 | 1 | 0 | Harry asks why people are overly optimistic about AGI. Richard highlights historical technological S-curves, citing flight speed plateaus and space shuttle return methods. | |
| Plateaus in AI and the Sequential Limit of Language | 3 | 5 | 1 | 1 | Richard explains that language has cognitive limits because human perception is sequential, preventing language from becoming superhuman in the way images have plateaued. | |
| Unequal Distribution of Technology and Physical Bottlenecks | 3 | 5 | 1 | 1 | Harry highlights the contrast between technological plateaus and global public awareness. Richard notes that unconstrained physical tasks become expensive economic bottlenecks. | |
| Managing the Speed of Job Displacement | 3 | 4 | 1 | 2 | Harry asks if compressed timelines for job displacement warrant concern. Richard compares the transition to 19th-century agricultural shifts and advocates for social safety nets. | |
| The Three Biggest Barriers to AGI | 2 | 5 | 2 | 0 | Richard argues that next-token prediction is not true AGI, pointing out that general intelligence requires intrinsic goal-setting which commercial incentives currently ignore. | |
| Quick Fire Round and Vision for You.com | 2 | 3 | 2 | 1 | In a rapid-fire round, Richard explains why he refused to sign Elon Musk's pause petition, using a Tesla self-driving update analogy. |
Statements from this episode (37)
Socher: True intelligent existence requires setting independent goals
“I think an intelligent existence probably needs to have some of its own goals and like some of its own mind of what it might want to do.”
Socher: You.com displays live stock tickers instead of LLM-generated estimates
“Like new.com has, for instance, when you ask for a stock price, we don't just make up a bunch of numbers like an LL would. We actually show you a stock ticker and then you can see the actual price, right?”
Socher: Scientific research is shaped by egos, pettiness, and politics
“Some people think research is just this purely truth seeking endeavor, but there are egos, there are people, and you can watch Oppenheimer as one of the, basically a movie about academic pettiness, you think into politics. And so that, that happens also in, in…”
Socher: OpenAI cited his 2018 multi-task NLP research in early GPT paper
“When in 2018, we invented problems engineering and we got this single model to work for all of these different NLP tasks. You know, it inspired some folks at OpenAI. They cited our paper in their first GPT open papers and cited our research there.”
Socher: AI advances faster by expanding unified models than building separate ones
“Kind of like imagine Wikipedia and everyone just keeps adding to Wikipedia and keeps making one dictionary better rather than everyone who wants to build a dictionary just starts their own dictionary company and then builds it from scratch. It just doesn't mak…”
Socher: DecaNLP paper was initially rejected by academic reviewers
“Actually the paper, when we submitted this prompt engineering paper, it got rejected.”
Socher: The human brain proves a single unified NLP model is possible
“The brain, like we don't replace our brain. If we do some math and then we ask like, you know, a sentiment question and then we ask a translation question, we use the same brain. So there is an existence proof for a single model. For all of NLP”
Socher: Incumbents hold a data advantage over startups in AI customer support
“If you want to answer customer emails automatically, it's very helpful if you're Salesforce and you already have all those emails, you already have labels that people, you know, mark that this knowledge-based article answered this email or answered this questi…”
LLMs allow small startups to rapidly build 80% AI product MVPs
“Because of these large language models and foundational models we can have a general sense of understanding natural language, and you can, as a small startup nowadays, build in like, 80% solution very quickly. And then you layer more and more specific data for…”
Socher: Startup moats come from distribution and partnerships, not AI models
“Turns out you can have moat other than your backend AI model, right? It's distribution. It's partnerships. It's your sales funnel processes and so on.”
Socher: LLMs are reasoning engines, not factual knowledge stores
“And if anything, you can think of these LMs as like reasoning engines, but you still need to feed them with the right facts and information. So that they can reason over the right things rather than just sort of reason what they remember.”
Socher: Calling You.com a thin LLM wrapper is completely false
“But it's something that, you know, some VCs don't appreciate and understand the complexity of, and then they say, oh, like u.com is just a thin wrapper around a large language model, which is very far from the truth.”
Socher: AI safety guardrails have overshot by censoring benign creative writing
“And I think some cases we even overshot a little bit. You know, there are some folks who say, oh, you cannot tell any dirty jokes anymore, or like a murder mystery is like unethical to write.”
Socher: You.com updates its production models bi-weekly
“Like we update our model every other week.”
Socher: GPU incompatibility prevents exploration of non-transformer AI models
“Transformers are mostly amazing because they can be trained on GPUs. There are lots of other models that we're currently not exploring because they're not trainable on GPUs.”
Socher: Base LLMs will become commoditized like databases
“And I think it won't matter that much which database you use, just like it won't matter that much, which LM you use, but it matters what you do with it, how you tune it what kind of training data you add onto it to fine tune it.”
Socher: LLMs are inherently ill-suited for numerical financial forecasting
“Like an LM won't be as good in doing a financial forecast because that's not what they're trained on. Like LM's help a ton for all things, natural language, because they understand so much about natural language and will have so much world knowledge, but that …”
Socher: Open-source GPT-4 equivalent model will launch before end of 2023
“I predicted that we'll have a GBD four equivalent model before the end of the year. That's open source. Of course, GBD four keeps getting better and better. So my prediction was for the version we had like a few months ago,”
Socher: You.com runs its own language model in production
“Certainly like u.com we partner with different providers, but we also have our own LM running in production already.”
Socher: Governments struggle to fund risky science due to taxpayer backlash
“It's very hard for governments to fund uncertain research and science projects because when it doesn't work out, you know, like some taxpayers will for sure complain.”
Socher: You.com's open platform struggled due to lack of user scale
“Now we've launched this open platform last year, but to be honest, we haven't had A ton of really amazing apps being added to this platform because we just don't have hundreds of millions of users, and so if your, you know, app, like, lets you book a kayaking …”
Socher: Content creators will be left out if closed search engines win
“If we don't win and others win, then they'll be actually fully left out.”
Socher: Google deranked You.com across all sites, causing immediate traffic drop
“We just got deranked in all our sites from Google and saw like a drop right away.”
Socher: Bing and Google copied You.com's AI chat features
“Bing and Google copy, you know, what we have launched last late last year with you chat. And you know, they may have worked on it before, like some people, you know, claim, oh, for, I mean, we've also met the prompt engineering in 2018, but like, they've copie…”
Socher: Google Earns $500 Million Daily From Main Search Ads
“They make five hundred million dollars a day with privacy invading advertisements on that page.”
Socher: Google Will Transition Core Search to AI Chat Slowly
“And so there is still some innovators dilemma that will not make them change their main experience overnight. It's been very carefully tuned. Every shade of blue has been AB tested, you know, to death. And like you, it's very hard to make a massive change and …”
Socher: Prompt engineering was invented at Salesforce Research
“When we invented prompt engineering, we actually did it at Salesforce research when I was chief scientist there.”
Socher: You.com launched the world's first search-integrated LLM
“We've launched uChap and had the first LM with a search backend and citations and web links and so on in the search context which we launched last December before anyone else in the world”
Socher: AI image generation capabilities have effectively maxed out
“And that is an image generation. It's photorealistic now. Look at some of these mid journey images. You can just say, I want this exact thing. And it'll just give you a photorealistic image of the scene you just described. Where do you go from there? You can't…”
Socher: Language AI faces a hard limit because language is human-bounded
“Language being a human construct and being how humans communicate thought and ideas, it has a limit on how superhuman it can be. It doesn't make sense to have superhuman language because humans couldn't understand it anymore to some degree.”
Socher: AI automation scales directly with how digitized and data-rich a job is
“The more digital they are, the more data there is about a job the more that job is likely to go to be automated and improved massively in its efficiency.”
Socher: Full self-driving in complex, unconstrained environments won't work for a long time
“Full self-driving on off-road, dirt road, like small roads, nighttime fog, like all of these things like won't work for a long time”
Socher: Physical jobs will become expensive bottlenecks constraining AI GDP growth
“The tasks that are physical are getting more and more expensive and they're going to become the new bottlenecks, right? So your carpenter, the people like who built your house all of those kinds of jobs are going to be more and more expensive. And then they're…”
Socher: Next-token prediction alone does not constitute general intelligence
“I find it hard to call something Artificially super intelligent or generally intelligent. If it, all it does is predict the next tokens and you say, oh, predict this next token. It'll predict that next token. I think an intelligent existence probably needs to …”
Socher: Commercial incentives prevent developers from building goal-setting AI
“Because companies need to make money and governments want to have a productive economy, no one is working on AI, just doing whatever it wants to do because that doesn't make any money. So no, one's working on AI setting its own goals.”
Socher: AI community should focus on current risks over existential sci-fi scenarios
“To have more folks talk to each other about real risks and let, you know, a small set of folks continue to think about existential risks, but not scare people so much with very interesting sci-fi and just general fiction scenarios that would make fun action mo…”