TOPIC
Foundation Models
A subject the collection covers, read through 29 synthesis notes.
View as
How fast is AI inference actually becoming more efficient?
Disentangling pure algorithmic progress from hardware improvements and market competition helps isolate the true rate of AI efficiency gains. This matters for understanding whether progress is driven by breakthroughs or just cheaper chips.
Is benchmark progress outpacing the cost to achieve it? Are AI feedback loops strong enough to sustain recursive self-improvement? Does inference cost remain a persistent economic burden? How do efficiency gains and new AI capabilities balance each other? How fast are frontier AI training costs actually rising?
Benchmark price-performance improves 5x to 10x yearly while running costs rise 3x to 18x AI feedback loops are strengthening but not yet self-sustaining today Inference cost per capability level falls roughly 5–10x yearly AI efficiency gains and energy-intensive capabilities create offsetting trends Frontier AI training costs grew 2.4× yearly since 2016
Can we distinguish real AI productivity gains from statistical measurement artifacts? Does real world AI cost drop as fast as benchmark performance prices? Why do AI performance prices fall fastest right after reaching state of the art? What efficiency gains could slow the growth in AI training costs? Can inference prices fall faster than AI usage grows? How do energy efficiency measurements compare to price-based algorithmic efficiency estimates? How much of AI capability gains come from spending versus algorithmic efficiency? How much inference efficiency improves per year after controlling for hardware advances? Why does high-end benchmark performance decline in price faster than low-end? Could AI inference capacity become scarcer than electricity for data centers? How does algorithmic efficiency differ from energy efficiency per task?
Is benchmark progress outpacing the cost to achieve it?
While benchmark scores improve dramatically year over year, the computational cost of running frontier models is rising even faster. Does this mean progress is illusory, or driven by spending rather than genuine capability gains?
Do automated benchmarks hide what frontier AI systems can really do? How fast is AI inference actually becoming more efficient? Does inference cost remain a persistent economic burden? How fast is AI capability becoming cheaper to produce? Do AI performance prices fall faster when capabilities debut?
Open-world evaluations reveal benchmark distortions in frontier AI capability Algorithmic efficiency in AI inference improves around 3x per year Inference cost per capability level falls roughly 5–10x yearly AI performance cost fell 47% quarterly since 2023, outpacing all historical technologies AI performance prices fall fastest right after reaching state of the art
How do different financing terms affect total cost of ownership? How much of frontier AI cost growth comes from chips versus skilled staff? What data-center infrastructure costs are excluded from frontier model training estimates? What time horizon does the two-year performance price slowdown suggest? Why do AI performance prices fall fastest right after reaching state of the art? What is driving the price increase in high-bandwidth memory supplies? Will memory's share of chip costs continue rising through 2026? Will the performance gap between open and frontier models widen as frontier models scale larger? Does hardware scaling alone explain why open models stay within months of frontier capability? How stable is the current 6 to 12 month lag between open and frontier models? How do energy efficiency measurements compare to price-based algorithmic efficiency estimates? Why does cost belong on the optimization axis alongside accuracy?See all 17 inquiring lines on this note →
What specific phrases reveal which model wrote a response?
Can simple word frequency analysis identify ChatGPT versus Claude outputs as reliably as complex neural classifiers? This explores whether model identity reduces to concrete linguistic habits or requires deeper pattern detection.
Today's chat models are reliably fingerprintable by word choice alone Imitation training captures style but closes no capability gap LLMs excel at explicit patterns and fail at inferred structural depth AI slop decomposes into three measurable dimensions with testable proxies Synthesis layers organize research around thematic maps and extraction workflows.
Do different AI models actually produce diverse outputs?
Explores whether using multiple different language models together creates genuine diversity or whether shared training and alignment cause them to converge on similar answers despite independence.
Does polished AI output trick audiences into trusting it? Why do LLMs generate novel ideas from narrow ranges? Why do preference models favor surface features over substance? Why do multi-agent LLM systems converge without genuine deliberation? Does model diversity actually reduce validator agreement failures?
Polished AI artifacts exploit professional appearance to simulate expertise LLM research ideation collapses into narrow clusters despite high novelty Preference models systematically favor five surface features humans reject Multi-agent LLM systems fail through silent agreement in over 60 percent of iterations Model family diversity alone may not restore validator independence
What happens when models train on AI-generated content recursively? Why do different AI models generate similar outputs independently? Why do different language models independently produce similar outputs? Can AI output be genuinely novel or only at the margins? Do AI-generated posts crowd out human voices without any coordination or intent? Why do multiple language models independently produce similar outputs in influence campaigns? Why does RLHF alignment reduce the diversity of viewpoints in AI output? What happens to solidarity and community signaling when AI smooths out voice differences? Can few-shot examples narrow generative diversity in creative tasks? Why do sigmoid conflict curves look the same across different language models? Does optimizing directly for semantic diversity improve both reasoning quality and exploration? Does alignment training create bidirectional instruction and response mappings?See all 142 inquiring lines on this note →
Will AI inference demand outpace supply capacity soon?
Emberson and Sevilla project token demand growing roughly 10× per year against inference supply expanding only 3.4× per year. The question is whether this gap will create a near-term compute bottleneck, especially for long-context and agentic workloads.
Inference compute will far exceed training as agents proliferate Meta's token leaderboard gamifies consumption without measuring output quality Frontier AI training costs grew 2.4× yearly since 2016 IEA projects data center electricity demand doubling to 950 TWh by 2030 Inference cost per capability level falls roughly 5–10x yearly
Does annualized capex or operating expense drive inference pricing? How much of token demand growth comes from context length versus user volume? What is driving the price increase in high-bandwidth memory supplies? Can inference prices fall faster than AI usage grows? Can inference demand itself change based on model price and speed? How much inference efficiency improves per year after controlling for hardware advances? What could slow or accelerate the IEA's projected data center growth? Why are energy supply chains becoming bottlenecks for data center expansion? What specific AI improvements drove the observed seven-month doubling rate?
What actually drives AI data center costs?
Does energy dominate AI infrastructure spending, or do hardware and capital costs dwarf operating expenses? Understanding cost drivers matters for predicting AI scaling economics and inference pricing.
Persistent agents shift economic unit from token to artifact Harness swaps move cost, not accuracy, beyond rerun noise AI agents show minimal real progress on NanoGPT after accounting for noise Frontier AI training costs grew 2.4× yearly since 2016 IEA projects data center electricity demand doubling to 950 TWh by 2030
How sensitive is data center cost to hardware depreciation assumptions? Does annualized capex or operating expense drive inference pricing? What margin do cloud providers add on top of infrastructure costs? How do different financing terms affect total cost of ownership? How is Meta's internal token usage driving AI infrastructure decisions? Why do AI labs lose more money than they earn from revenue? How much of frontier AI cost growth comes from chips versus skilled staff? Do R&D labor costs scale with computing power the same way hardware costs do? What data-center infrastructure costs are excluded from frontier model training estimates? Does competitive pricing pressure explain most AI cost decline patterns? How much do typical users actually save from declining AI inference costs? Does real world AI cost drop as fast as benchmark performance prices?See all 42 inquiring lines on this note →
How long do leading AI chips stay in frontier training?
Frontier AI training depends on specialized hardware. Understanding how long chips remain competitive for the largest training runs matters for hardware amortization planning and predicting when designs become obsolete.
Frontier AI training consolidated onto fewer chip designs after V100 Pretraining loss predicts post-RL performance in chess models Open models on consumer GPUs match frontier performance within 6 to 12 months Open models gain LM Arena Elo roughly 56 percent faster than frontier models R&D staff make up 29 to 49 percent of frontier AI training costs
Why did frontier training concentrate onto fewer chip designs after the V100? Do custom ASICs and new TPU generations extend chip lifespan beyond four years? How long do leading AI chip designs stay in use for frontier training? What drives consolidation of frontier training onto specific hardware vendors? Did non-frontier AI training spread across more chip designs than frontier training? Why did frontier training consolidate around specific chip designs after V100? How much of frontier AI cost growth comes from chips versus skilled staff? How much of frontier AI training costs come from staff versus hardware?
Where did OpenAI actually spend its 2024 compute budget?
Explores how OpenAI's $5 billion R&D compute spending was distributed across released models, experimental runs, and unreleased work. Understanding this breakdown clarifies whether major models like GPT-4o recouped their development costs.
R&D staff make up 29 to 49 percent of frontier AI training costs GPT-5 likely failed to recoup its R&D costs in four months AI labs spend 2-3x their revenue with divergent profitability timelines Frontier AI training costs grew 2.4× yearly since 2016 Compute costs dominate AI company spending across three labs
How far behind closed models are open-weight models?
Epoch AI's capability index measures the time lag between frontier closed models and the most capable open-weight alternatives, examining whether open-source development keeps pace with proprietary labs.
Frontier exams discriminate temporarily but miss autonomous research ability Capability gains are spiky and scaffolding-driven, not steady or weight-dependent alone AI cyber task autonomy is doubling every few months and accelerating Open models gain LM Arena Elo roughly 56 percent faster than frontier models Open models on consumer GPUs match frontier performance within 6 to 12 months
How fast is AI capability becoming cheaper to produce?
Epoch AI tracked whether the cost of achieving fixed AI performance levels is falling, and if so, how that decline compares to other transformative technologies throughout history.
AI performance prices fall fastest right after reaching state of the art AI agents show minimal real progress on NanoGPT after accounting for noise AI agents outperform humans at 2 hours but humans gain ground with more time Benchmark price-performance improves 5x to 10x yearly while running costs rise 3x to 18x Frontier AI training costs grew 2.4× yearly since 2016
What time horizon does the two-year performance price slowdown suggest? Does real world AI cost drop as fast as benchmark performance prices? Why do AI performance prices fall fastest right after reaching state of the art? What timeline do AI labs need to recoup model development costs before obsolescence? How quickly will AI service costs fall enough to trigger widespread automation? Why does high-end benchmark performance decline in price faster than low-end?
Did AI chip designs consolidate after the V100?
Exploring whether frontier AI training shifted from spreading across diverse chip designs to concentrating on fewer, heavily-reused architectures around the time of NVIDIA's V100 release.
Leading AI chips serve frontier training for roughly four years Pretraining loss predicts post-RL performance in chess models Open models on consumer GPUs match frontier performance within 6 to 12 months R&D staff make up 29 to 49 percent of frontier AI training costs Open models gain LM Arena Elo roughly 56 percent faster than frontier models
Why did frontier training concentrate onto fewer chip designs after the V100? Do custom ASICs and new TPU generations extend chip lifespan beyond four years? How long do leading AI chip designs stay in use for frontier training? What drives consolidation of frontier training onto specific hardware vendors? Did non-frontier AI training spread across more chip designs than frontier training? Why did frontier training consolidate around specific chip designs after V100? Will custom ASICs and new architectures reverse the chip design consolidation trend? How much of frontier AI cost growth comes from chips versus skilled staff? Which spending categories drive the rising cost of frontier AI training? How much of frontier AI training costs come from staff versus hardware?
How much do AI staff costs drive frontier model training?
Frontier AI models require massive spending on both chips and skilled R&D labor. Understanding the breakdown between hardware and human costs helps clarify what makes cutting-edge AI development so expensive.
Frontier AI training costs grew 2.4× yearly since 2016 Data moats weaken as collection costs rise and marginal value falls Compute costs dominate AI company spending across three labs Most of OpenAI's 2024 R&D compute went to experimental and unreleased work Servers, not energy, dominate AI data center costs
Why did frontier training concentrate onto fewer chip designs after the V100? How long do leading AI chip designs stay in use for frontier training? What drives consolidation of frontier training onto specific hardware vendors? How much of frontier AI cost growth comes from chips versus skilled staff? Do R&D labor costs scale with computing power the same way hardware costs do? How does staff cost share differ between individual training runs and whole company budgets? Which spending categories drive the rising cost of frontier AI training? How much of frontier AI training costs come from staff versus hardware?
Do AI performance prices fall faster when capabilities debut?
Explores whether newly state-of-the-art AI capabilities command premium pricing that erodes over time, and how quickly that erosion happens across different benchmarks.
AI performance cost fell 47% quarterly since 2023, outpacing all historical technologies AI agents show minimal real progress on NanoGPT after accounting for noise AI agents outperform humans at 2 hours but humans gain ground with more time Benchmark price-performance improves 5x to 10x yearly while running costs rise 3x to 18x Algorithmic efficiency in AI inference improves around 3x per year
How sensitive is data center cost to hardware depreciation assumptions? Does annualized capex or operating expense drive inference pricing? What margin do cloud providers add on top of infrastructure costs? Why do AI labs lose more money than they earn from revenue? How are enterprises shifting their spending on AI capabilities? How much of frontier AI cost growth comes from chips versus skilled staff? Do R&D labor costs scale with computing power the same way hardware costs do? Does competitive pricing pressure explain most AI cost decline patterns? What time horizon does the two-year performance price slowdown suggest? Do AI capabilities mature on a predictable price decay schedule? How universal is the state-of-the-art premium effect across AI benchmarks? Does real world AI cost drop as fast as benchmark performance prices?See all 28 inquiring lines on this note →
Do compute costs really dominate AI company budgets?
Explores whether research and inference compute truly represent the largest expense category at major AI labs, and how this compares to staffing and other costs.
OpenAI's research agents now log more workdays than human researchers Enterprise AI budgets shifted from innovation to core recurring spending Organizational practices, not model quality, determine AI's business impact AI labs spend 2-3x their revenue with divergent profitability timelines R&D staff make up 29 to 49 percent of frontier AI training costs
Does annualized capex or operating expense drive inference pricing? What margin do cloud providers add on top of infrastructure costs? Why do AI labs lose more money than they earn from revenue? How are enterprises shifting their spending on AI capabilities? How much of frontier AI cost growth comes from chips versus skilled staff? Do R&D labor costs scale with computing power the same way hardware costs do? What data-center infrastructure costs are excluded from frontier model training estimates? Does competitive pricing pressure explain most AI cost decline patterns? How much do typical users actually save from declining AI inference costs? Does real world AI cost drop as fast as benchmark performance prices? Why do AI labs face shorter repayment windows than traditional software companies? How do gross margins and operating margins differ for AI model providers?See all 26 inquiring lines on this note →
Is memory becoming the dominant cost in AI chips?
Epoch AI's cost model tracks whether high-bandwidth memory has overtaken other components to dominate AI chip spending, and what that shift means for hyperscaler budgets and chip design priorities.
Inference compute will far exceed training as agents proliferate Enterprise AI budgets shifted from innovation to core recurring spending Frontier AI training costs grew 2.4× yearly since 2016 Algorithmic efficiency in AI inference improves around 3x per year R&D staff make up 29 to 49 percent of frontier AI training costs
Do custom ASICs and new TPU generations extend chip lifespan beyond four years? Will custom ASICs and new architectures reverse the chip design consolidation trend? How are hyperscalers adjusting capex budgets in response to memory cost inflation? What is driving the price increase in high-bandwidth memory supplies? Will memory's share of chip costs continue rising through 2026? Does inference demand for memory-hungry accelerators explain the shift toward memory-dominant bills of materials? Which spending categories drive the rising cost of frontier AI training? How much does electricity actually cost in a large AI data center? When will energy-intensive AI tasks become the majority of AI workloads?
How quickly do open models catch up to frontier AI?
Epoch AI investigates whether open-weight models small enough to run on consumer GPUs match frontier AI performance. The question matters because it determines how soon most people can access near-state-of-the-art capabilities locally.
Small language models suffice for most agentic subtasks cost-effectively Mobile LLMs are constrained by hardware limits not quality preference Frontier exams discriminate temporarily but miss autonomous research ability Open models gain LM Arena Elo roughly 56 percent faster than frontier models Open-weight models lag closed frontier models by four months
Are open models closing on frontier AI faster than expected?
Epoch AI's LM Arena Elo analysis tracks whether consumer-GPU-sized open models are gaining capability faster than frontier closed models. This matters because it suggests whether the gap to frontier AI is narrowing or holding steady.
Open models on consumer GPUs match frontier performance within 6 to 12 months Frontier exams discriminate temporarily but miss autonomous research ability Open-world evaluations reveal benchmark distortions in frontier AI capability Open-weight models lag closed frontier models by four months Benchmark price-performance improves 5x to 10x yearly while running costs rise 3x to 18x
Will the performance gap between open and frontier models widen as frontier models scale larger? Does hardware scaling alone explain why open models stay within months of frontier capability? How stable is the current 6 to 12 month lag between open and frontier models? Can model release timing distort the measured slope of Elo progress? Why are consumer GPU open models advancing faster on LM Arena metrics? Why do open-weight models capture significant share if closed models score higher on benchmarks? Why do closed-weight models show faster price declines than open-weight ones? Why does high-end benchmark performance decline in price faster than low-end? How quickly do open models close the performance gap after closed model releases?
Do AI labs spend far more than they earn?
Epoch AI examined whether leading AI companies operate at massive losses relative to revenue. This matters because it determines whether the current AI business model is sustainable or requires fundamental changes.
Compute costs dominate AI company spending across three labs Organizational practices, not model quality, determine AI's business impact AI productivity claims need independent verification before counting as real savings Enterprise AI budgets shifted from innovation to core recurring spending Most of OpenAI's 2024 R&D compute went to experimental and unreleased work
How fast are frontier AI training costs actually rising?
Researchers at Epoch AI estimated the rate at which the most expensive AI model training runs are becoming costlier. Understanding this trend matters because it shapes who can afford to build the most capable AI systems.
How much do AI staff costs drive frontier model training? Does data really create lasting competitive advantage for startups? Are AI feedback loops strong enough to sustain recursive self-improvement? Is benchmark progress outpacing the cost to achieve it? Where did OpenAI actually spend its 2024 compute budget?
R&D staff make up 29 to 49 percent of frontier AI training costs Data moats weaken as collection costs rise and marginal value falls AI feedback loops are strengthening but not yet self-sustaining today Benchmark price-performance improves 5x to 10x yearly while running costs rise 3x to 18x Most of OpenAI's 2024 R&D compute went to experimental and unreleased work
Does annualized capex or operating expense drive inference pricing? How long do leading AI chip designs stay in use for frontier training? What drives consolidation of frontier training onto specific hardware vendors? How much of frontier AI cost growth comes from chips versus skilled staff? How does staff cost share differ between individual training runs and whole company budgets? Why do AI performance prices fall fastest right after reaching state of the art? Which spending categories drive the rising cost of frontier AI training? What timeline do AI labs need to recoup model development costs before obsolescence? What efficiency gains could slow the growth in AI training costs? How much of frontier AI training costs come from staff versus hardware?
Did GPT-5 earn back its research and development costs?
Epoch AI reconstructed OpenAI's finances to test whether GPT-5's four-month revenue was enough to recoup its R&D spending. The question matters because it separates whether a model finds paying customers from whether its maker actually profits.
Enterprise AI success depends on deployment approach, not model quality Nearly half of companies now capture meaningful AI value Organizational practices, not model quality, determine AI's business impact Most of OpenAI's 2024 R&D compute went to experimental and unreleased work Compute costs dominate AI company spending across three labs
How do efficiency gains and new AI capabilities balance each other?
AI energy efficiency is improving dramatically, but new capabilities like video generation and agentic tasks demand far more power per query. Understanding how these forces net out is crucial for predicting AI's actual electricity footprint.
IEA projects data center electricity demand doubling to 950 TWh by 2030 Algorithmic efficiency in AI inference improves around 3x per year AI performance cost fell 47% quarterly since 2023, outpacing all historical technologies Frontier AI training costs grew 2.4× yearly since 2016 Token demand is growing faster than inference supply capacity
How much more energy does video generation consume than text queries? How much does electricity actually cost in a large AI data center? Does an AI productivity boost automatically increase electricity consumption? When will energy-intensive AI tasks become the majority of AI workloads? Why are energy supply chains becoming bottlenecks for data center expansion? How does algorithmic efficiency differ from energy efficiency per task?
How fast will data center electricity demand actually grow?
The IEA projects data center power consumption will roughly double by 2030, driven largely by AI infrastructure buildout. Understanding whether this central projection will hold depends on uncertainties in efficiency, adoption rates, and financing conditions.
Servers, not energy, dominate AI data center costs Token demand is growing faster than inference supply capacity AI efficiency gains and energy-intensive capabilities create offsetting trends Frontier AI training costs grew 2.4× yearly since 2016 High-bandwidth memory grew from 52% to 63% of AI chip spending
Can deep learning theory unify around training dynamics?
Is learning mechanics—focused on average-case predictions and training dynamics rather than worst-case bounds—the emerging framework that finally unifies fragmented deep learning theory?
Language models show three hierarchical tiers of understanding with distinct mechanisms Marr's three levels provide structured methodology for LLM interpretability Human-parseable theory is essential for AI safety oversight regardless of AI self-understanding Shannon and Kolmogorov measures fail to value data for bounded learners Epiplexity measures learnable structure independent of computational resources needed
What stability techniques prevent collapse in policy-critic adversarial training? How does mechanistic interpretability complement learning mechanics in explaining deep learning? Why should deep learning theory prioritize average-case over worst-case analysis? Which hyperparameter theories best explain universal behaviors across neural networks? What solvable idealized settings reveal fundamental phenomena in realistic deep learning? How do classical mechanics and statistical mechanics provide methodological templates for learning theory? How does the Learning Law explain why all examples should contribute equally? Why do optimal learning dynamics improve scaling law coefficients specifically? What distinguishes surface mechanisms from the training regimes that produce them? How do learning dynamics on one example shift predictions on other responses?
Can we identify which LLM wrote a given text?
Do language models leave distinctive fingerprints in their outputs that survive rewriting and translation? This explores whether model-specific word choices are robust enough to trace a text's origin.
Can humans detect AI text if machines can measure it? Can imitating ChatGPT fool evaluators into thinking models improved? What specific phrases reveal which model wrote a response? Why do newer AI models diverge further from human writing patterns? Can human judges detect measurable differences in AI text?
AI text diverges measurably from human text but remains imperceptible to judges Imitation training captures style but closes no capability gap ChatGPT and Claude use distinct opening phrases and markdown formatting habits Newer AI models diverge from human lexical patterns while becoming harder to detect LLM text shows measurable lexical differences humans cannot perceive
Why do closed AI models dominate despite costing six times more?
Closed models command 80% of usage and 96% of revenue even though open alternatives reach 90% of their performance at a fraction of the cost. What explains this apparent market disconnect?
Open-weight models serve over 30% of tokens despite closed model dominance Price alone poorly predicts LLM usage across OpenRouter's market Small model deployments break even in months, large ones in years Cascading through cheaper models matches GPT-4 accuracy with 98% cost reduction Leading LLM routers converge in performance and some lag simple baselines
When does self hosting an LLM beat commercial API costs?
This analysis explores the financial payback period for organizations deploying open-source LLMs on their own hardware versus paying per-token to commercial API providers, across different model sizes and usage volumes.
Are enterprises abandoning custom AI apps for third-party solutions? Can small language models handle most agent tasks? Does LLM price alone determine which model users choose? Why do closed AI models dominate despite costing six times more? Should benchmarks report token counts instead of frozen costs?
Enterprises are shifting from building to buying AI applications Small language models suffice for most agentic subtasks cost-effectively Price alone poorly predicts LLM usage across OpenRouter's market Closed models dominate through brand trust and information gaps, not performance Benchmarks must report token counts so downstream costs stay current
How sensitive is data center cost to hardware depreciation assumptions? What margin do cloud providers add on top of infrastructure costs? How do different financing terms affect total cost of ownership? How do latency and reliability trade off against price in real LLM adoption decisions? How does data residency constraint change the economics of self hosted LLMs? Which open source models meet the threshold for viable on-premise deployment?
How much do open-weight models actually get used in practice?
Real-world inference routing data can reveal whether open models capture meaningful market share despite proprietary model advantages. Understanding actual usage patterns matters for predicting the future of AI infrastructure.
Small language models suffice for most agentic subtasks cost-effectively Price alone poorly predicts LLM usage across OpenRouter's market Closed models dominate through brand trust and information gaps, not performance Leading LLM routers converge in performance and some lag simple baselines LLM routers decompose into five comparable components
Do public benchmarks systematically overstate open model capabilities? How much capability remains hidden in unreleased proprietary models? Why are consumer GPU open models advancing faster on LM Arena metrics? Why do users pick cheaper open models when closed models perform better? Do open-weight models dominate particular task categories over proprietary alternatives? Why do open-weight models capture significant share if closed models score higher on benchmarks? Why do closed-weight models show faster price declines than open-weight ones? Will open-weight and closed-weight algorithmic efficiency trends diverge in the future? Which open source models meet the threshold for viable on-premise deployment? How quickly do open models close the performance gap after closed model releases? Do geopolitical concerns about Chinese open models affect enterprise adoption decisions? Does curating smaller model pools improve router recall compared to scaling?See all 14 inquiring lines on this note →
Does LLM price alone determine which model users choose?
If cheaper models offer 10-30x cost savings, why don't they dominate adoption? This explores whether price is the primary driver of LLM selection or whether other factors like quality and reliability matter equally.
Small language models suffice for most agentic subtasks cost-effectively Open-weight models serve over 30% of tokens despite closed model dominance Closed models dominate through brand trust and information gaps, not performance Small model deployments break even in months, large ones in years Cascading through cheaper models matches GPT-4 accuracy with 98% cost reduction
How do different financing terms affect total cost of ownership? How much does price alone explain differences in LLM usage patterns? Why do users pick cheaper open models when closed models perform better? How do latency and reliability trade off against price in real LLM adoption decisions? What evidence shows users genuinely prioritize factors beyond cost when selecting models?
Can humans understand deep learning before AI does?
Explores whether investing in human-parseable deep learning theory remains valuable even if AI systems eventually develop their own self-understanding. Centers on why this matters for safety oversight.
Can deep learning theory unify around training dynamics? Can we monitor AI reasoning without destroying what makes it readable? Does incremental AI replacement erode human influence over society? Can global standards pace frontier AI as much as alignment research? What happens to social order when AI removes ritual constraints?
Learning mechanics unifies deep learning theory through training dynamics Optimizing reasoning traces for safety produces obfuscation instead of alignment Incremental AI adoption degrades implicit alignment by replacing human labor dependency. OpenAI argues global standards pace frontier AI as much as alignment research AI conversation lacks the ritual machinery that sustains human social order.
Why do human-designed neural architectures eventually get replaced by learned ones? What makes a neural network circuit actually interpretable to humans? What does a human-parseable framework for deep learning look like? Why does human-AI collaboration preserve safety compared to autonomous self-improvement?