AI Vibe Check: The Actual Bottleneck In Research, SSI’s Mystique, & Spicy 2026 Predictions

Ari reflects on attending his 11th NeurIPS, noting growth from 2,000 to 30,000 attendees since 2015. Despite feeling 'late' in 2015, he now believes we're still extremely early in AI. However, NeurIPS has become more about PR than sharing work, as labs like Meta and Google now prohibit publishing meaningful research, creating a 'Goldilocks zone' where papers must be good enough to pass review but not important enough to matter for models.

• •NeurIPS grew 15x from 2,000 (2015) to 30,000 attendees (2024), with 8,000 increase from previous year alone
• •Major labs now have policies preventing submission of meaningful work - only allowing papers in narrow 'Goldilocks zone'
• •Researchers remove large-scale results from papers to game the publication system while keeping proprietary advantages
• •Despite massive growth, we're still 'extremely early' in AI - similar to how 2015 felt late but wasn't

Rob argues models are clearly plateauing, citing decreased incremental improvements since GPT-4 and fundamental limitations in continual learning and sample efficiency. Ari counters that we're in an 'LLM bubble' not an AI bubble - while consumer LLMs may plateau, video models and other modalities show no signs of slowing. The community's myopic focus on language models misses breakthroughs happening elsewhere in AI.

• •GPT-1 to GPT-4 showed breathtaking capability leaps; post-GPT-4 improvements have significantly decreased
• •Fundamental problems like continual learning and sample efficiency remain unsolved in current paradigm
• •Video models have shown dramatic improvements throughout 2024 with no plateau in sight
• •Consumer LLM plateau doesn't mean AI progress has stopped - just that one modality is maturing

Ari explains RL only works in a specific 'Goldilocks zone' where models understand enough that on-policy guesses have some chance of being correct, but not so much that sparse rewards don't teach anything. Coding models hit this zone, but many domains haven't. The challenge isn't just applying RL - it's getting models to the readiness point where RL can be effective, which requires significant pre-work in most enterprise domains.

• •RL requires models in specific capability zone: not too easy (no learning) or too hard (no signal)
• •Coding models reached this Goldilocks zone, enabling rapid RL improvements
• •Most enterprise domains need substantial work to reach RL-readiness, not just RL application itself
• •Training stages (pre-training, mid-training, post-training, RL) are artificially separated - should be viewed holistically

Ari argues that unlimited funding at major labs is a 'huge detriment' because constraints breed innovation. Chinese labs have achieved incredible results with far fewer resources than individual Western labs because they must innovate on efficiency. When you have infinite capital, throwing more data and compute at problems becomes easier than solving hard research problems around sample efficiency, even though we've known about diminishing returns from scaling laws since Kaplan's foundational work.

• •XAI alone likely has more funding and compute than all Chinese labs combined, yet China achieves remarkable results
• •Scaling laws explicitly showed diminishing returns (power law) - yet labs are surprised infinite scaling doesn't work
• •With unlimited capital, buying more data is faster than solving sample efficiency - the 'easier, more expensive path'
• •Chinese labs forced to innovate on file systems, curation, and efficiency due to resource constraints

Debating recursive self-improvement and AI researchers, Ari challenges whether the bottleneck is actually lack of ideas versus compute to test them. OpenAI likely already has great ideas from researchers - the constraint is GPU bandwidth and time to experiment. An AI researcher only accelerates progress if it has dramatically higher hit rates than human researchers, similar to how AI for drug discovery works because current hit rates are 0.0001%. The analogy to bio is apt but assumes research hit rates are similarly abysmal.

• •Bottleneck isn't writing code or generating ideas - it's GPU compute and time to test at scale
• •AI researcher only helps if hit rate is orders of magnitude better than human researchers collectively
• •Similar to bio: AI helps drug discovery because current hit rates are ~0.0001%, making even 1% transformative
• •OpenAI likely already has strong ideas in their researcher pool - execution bandwidth is the constraint

SSI maintains unprecedented secrecy - no phones allowed, employees can't discuss AI at all to avoid inadvertent reveals. Rob shares an anecdote where Ilya told an interviewer there are 'two words' that would explain SSI's approach, but refused to share them. Two explanations exist: they've genuinely kept the next big thing secret, or there's nothing to leak. Ilya's track record (correctly calling scaling in 2018) demands respect, but aggressive secrecy makes it impossible to evaluate.

• •SSI security exceeds all other labs: phones in bags, employees forbidden from any AI discussion publicly
• •Ilya told interviewer 'two words' would explain SSI but refused to share - creating mystique
• •Two possibilities: genuinely revolutionary secret breakthrough, or nothing substantial to leak
• •Ilya's 2018 scaling prediction (when most were skeptical) gives him credibility for bold bets

OpenAI's 'code red' signals the end of their aura of infallibility as Google demonstrates advantages in talent depth, compute resources, and crucially, cash position. Google has hundreds of billions on balance sheet while OpenAI burns cash at unprecedented rates - projecting $150B cumulative burn before profitability in 2029 (versus Uber's record $40B). This makes OpenAI completely dependent on capital markets, unlike Google which can self-fund indefinitely.

• •OpenAI's mystique of being 'one or two steps ahead' is fading as competitors catch up
• •Google has structural advantages: deeper talent bench, more compute, vertical integration with TPUs
• •OpenAI projects $150B cumulative burn before 2029 profitability - 3.75x Uber's record $40B
• •Complete dependence on capital markets creates existential vulnerability if investor sentiment shifts

Ari calls chip restrictions 'one of the stupidest things we've ever done,' arguing it stems from viewing China as a 'copycat factory' unable to innovate - a view true in the 1990s but completely false today. China is genuinely innovating in AI (DeepSeek papers are must-reads at Western labs), and restrictions likely accelerated their domestic chip development by 5-10 years. The CCP's ability to think in 10-40 year timeframes, superior to democracies with election cycles, makes this a catastrophic strategic error.

• •Restrictions assume China can't innovate - true in 1990s, completely false in 2020s (DeepSeek proves this)
• •Likely accelerated China's chip self-sufficiency by 5-10 years by forcing domestic development
• •Without restrictions, China would have continued buying NVIDIA chips indefinitely with no urgency to develop alternatives
• •CCP's 10-40 year planning horizon (vs. Western election cycles) makes long-term chip independence inevitable

Ari reports being 'more bearish' on Meta's superintelligence team after numerous departures with extremely short tenures and cultural challenges. The Llama 4 'purge' removed many exceptional researchers, representing organizational dysfunction where Meta pays premium for AI talent while simultaneously letting great researchers go. FAIR, once a beloved research org, is 'on its last legs' with researchers deeply unhappy and seeking exits. The shift from open source (which rehabilitated Zuck's image) to closed source removes key tailwinds.

• •Multiple MSI departures with very short tenures signal cultural and organizational problems
• •Llama 4 'purge' removed exceptional researchers - dysfunction of paying premium while losing talent
• •FAIR researchers at NeurIPS were 'deeply unhappy' and actively looking to exit
• •Shift from open source (which rehabilitated Zuck's reputation) to closed source removes narrative advantage

Rob predicts Sam Altman won't be OpenAI CEO by end of 2026, citing shifting narrative, his scattered focus (BCI, space, chips), and the need for 'straight-laced' leadership for IPO preparation - comparing to Travis/Dara transition at Uber. Ari predicts >50% chance the world's best model will be Chinese open source at least once in 2026, with >20% chance it remains best at year-end. Jacob predicts infrastructure will create more value than applications in 2026 after years of app dominance, as model stability finally enables real infra companies to emerge.

• •Sam's scattered focus (BCI company Merge Labs, space, chips, hardware) problematic as OpenAI faces headwinds
• •Hiring Fidji Simo as 'CEO of applications' suggests Sam delegating day-to-day operations
• •Chinese open source models likely to achieve clear superiority at least once in 2026 (>50% probability)
• •Infrastructure investing will outperform applications in 2026 as model plateau creates stable substrate for infra innovation