NDC Conferences · Published 2026-06-18

Keynote: After the AI Hype – What’s Real, and What’s Next - Richard Campbell - 2026

privateReview: unreviewedConfidence: high
artificial intelligencellmhype cyclegptAlphaFoldai-hype-cyclegartner-hype-cyclecapexcircular-financingvendor-consolidationAnthropomorphismscaling-laws
Open on YouTube ↗

Summary

Overview

  • Speaker: Richard Campbell
  • Channel: NDC Conferences
  • Main topic: Artificial Intelligence hype, history, financial cycles, and current reality in software development
  • Purpose: To contextualize the current generative AI boom within historical technology cycles, debunk hype, and provide developers and organizations with a realistic perspective on AI's true capabilities and limitations. Richard Campbell explores the history of artificial intelligence, drawing parallels between current generative AI hype cycles and past technology booms like the .com bubble. He discusses the evolution of neural networks, ImageNet, Siri, GPT-2 to GPT-4, OpenAI, Microsoft, DeepSeek, and AlphaFold, analyzing the economic sustainability, infrastructure investments, and societal impact of modern AI tools.

Topic Map

History and naming of AI

  • Explanation: Exploration of the origins of the term 'artificial intelligence' in the 1950s by scientists raising funds from the US military, and the recurring 'AI winters'.
  • Key claims:
    • AI is a terrible name coined in the 1950s by scientists seeking military funding.
    • Humans are hardwired to perceive intelligence and faces everywhere (pareidolia), making us susceptible to AI hype.
  • Examples:
    • Joseph Weizenbaum's 1966 program ELIZA as a rudimentary psychotherapist.
    • Science fiction films like 2001: A Space Odyssey, The Terminator, and Avengers: Age of Ultron shaping public perception.
  • Terminology:
    • Artificial Intelligence
    • AI Winter
    • ELIZA
    • Pareidolia
  • Why it matters: Understanding that AI's naming and perception have been influenced by marketing and science fiction for decades helps separate reality from hype.

The emergence of Neural Networks and ImageNet

  • Explanation: How deep learning, backpropagation, and the ImageNet competition transformed computer vision and speech recognition.
  • Key claims:
    • Geoffrey Hinton's work on deep neural nets and backpropagation laid the foundation for modern deep learning.
    • The ImageNet competition in 2012 used 14 million labeled images to dramatically improve visual recognition.
  • Examples:
    • Siri created via Apple buying voice recognition tech trained by SRI researchers.
    • Google's acquisition of Hinton's team to form Google Brain.
  • Terminology:
    • Neural Nets
    • Backpropagation
    • ImageNet
    • Siri
    • Google Brain
    • Generative Pre-trained Transformer
  • Why it matters: Neural networks solved hard pattern recognition problems once compute power and large datasets became available.

The OpenAI, Microsoft, and hyperscaler economics

  • Explanation: Examining the financial backing, partnerships, and computing resource investments driving companies like OpenAI, Microsoft, and Google.
  • Key claims:
    • OpenAI started as a public research lab but shifted to heavy commercial funding and capital needs.
    • Microsoft invested billions in OpenAI in exchange for Azure cloud commitments (self-dealing model).
  • Examples:
    • Microsoft investing $1 billion initially, followed by more billions.
    • Neural scaling laws driving the push for larger models and massive supercomputing clusters.
  • Terminology:
    • Hyperscalers
    • Neural Scaling Laws
    • GPT-2
    • GPT-3
    • GPT-4
    • GPT-4o
    • Copilot
    • Capex
  • Why it matters: The economic engine of modern AI requires unprecedented capital expenditure on data centers and GPUs, distorting macroeconomic figures.

The Gartner Hype Cycle and comparison to the .com boom

  • Explanation: Mapping current AI adoption against the Gartner hype cycle and historical technology bubbles like the .com boom and Netscape IPO.
  • Key claims:
    • AI has hit the peak of inflated expectations and is sliding toward the trough of disillusionment.
    • Data center infrastructure buildout resembles the overbuilding of the .com boom.
  • Examples:
    • Netscape IPO as the technology trigger of the .com boom.
    • Hyperscaler capex figures representing over half of S&P 500 value increases in 2024.
  • Terminology:
    • Gartner Hype Cycle
    • Technology Trigger
    • Peak of Inflated Expectations
    • Trough of Disillusionment
    • CapEx
  • Why it matters: Recognizing hype cycles helps investors and developers avoid over-leveraging on temporary market manias.

AlphaFold and scientific breakthroughs

  • Explanation: How DeepMind's AlphaFold solved protein folding, revolutionizing biology and medicine.
  • Key claims:
    • Protein folding is one of biology's most complex challenges, with 10^35 possible combinations.
    • AlphaFold achieved 90% accuracy by 2020, computing 200 million protein folds and publishing them for free.
  • Examples:
    • John Kendrew winning a Nobel Prize after a decade of studying myoglobin.
    • New malaria vaccines and antibiotics developed using AlphaFold data.
  • Terminology:
    • AlphaFold
    • DeepMind
    • Protein Folding
    • Amino Acids
    • Demis Hassabis
  • Why it matters: Demonstrates real, profound scientific utility of AI beyond chat interfaces and marketing hype.

Key Points

AI is software, not a person

  • Explanation: People tend to anthropomorphize software and LLMs, treating them like therapists or conscious entities due to human evolutionary tendencies.
  • Evidence: Users becoming emotionally attached to chatbots or attributing human sentience to systems like ChatGPT.
  • Practical implication: Maintain professional skepticism and remember that LLMs are advanced predictive text engines lacking common sense.

Massive capital expenditure and infrastructure strains

  • Explanation: Hyperscalers are spending hundreds of billions on data centers and GPUs, creating significant power and land demands.
  • Evidence: Nvidia, Microsoft, Google, and Amazon capex numbers distorting broader economic growth metrics.
  • Practical implication: Organizations must evaluate the actual ROI of AI infrastructure investments against cooling market enthusiasm.

Real vs Hype in software engineering

  • Explanation: While coding assistants like GitHub Copilot provide genuine productivity boosts, many ancillary AI products fail to deliver sustainable profit.
  • Evidence: Proliferation and disappearance of numerous small AI coding tools as major players consolidate.
  • Practical implication: Focus on fundamentals and proven tools rather than chasing every new AI product wave.

Frameworks, Models & Processes

Gartner Hype Cycle

  • How it works: Tracks a technology's maturity from trigger through inflated expectations, disillusionment, and eventual productivity plateau.
  • Components:
    • Technology Trigger
    • Peak of Inflated Expectations
    • Trough of Disillusionment
    • Slope of Enlightenment
    • Plateau of Productivity
  • When to use: Evaluating emerging technologies to determine whether market enthusiasm matches realistic utility.

Examples & Case Studies

DeepMind's AlphaFold predicted protein structures with high accuracy.

  • Illustrates: Real, high-value scientific application of AI solving previously intractable problems.
  • Lesson: AI delivers transformative results when applied to domains with massive structured data and rigorous validation.

Cambridge Analytica exploited Facebook data for political advertising during Brexit.

  • Illustrates: The risks of manipulative AI systems optimizing for user engagement and deception.
  • Lesson: Powerful recommendation and targeting software can be misused when user engagement is the primary metric.

Actionable Takeaways

  • Immediate:
    • Avoid anthropomorphizing LLMs; treat them as sophisticated software tools.
    • Use coding assistants like GitHub Copilot critically while maintaining personal accountability for code.
  • Strategic:
    • Prepare for the trough of disillusionment as AI hype cycles mature and investment returns face scrutiny.
    • Focus organizational strategy on proven productivity gains rather than speculative AI general intelligence promises.
  • Questions to investigate:
    • How sustainable are current hyperscaler capital expenditures on AI data centers?
    • What are the long-term regulatory and legal liability frameworks for AI-generated code and content?

Claims Worth Verifying

  • Elon Musk's promised $100 million investment in OpenAI turned out to be $16 million and 8 Teslas. (anecdotal financial claim)
  • More than half of the increase in value in the S&P 500 in 2024 came from seven tech companies. (macroeconomic market statistic)

Notable Quotes

"The term artificial intelligence, it's a terrible name. It's an old term. It's from the 1950s." (at 0:25) "Put down the software. Remind yourself that it is software. It's the price of the deception to do the fundraising." (at 68:32)

Compressed Summary

  • AI's name originates from 1950s military funding proposals.
  • Neural scaling laws and massive GPU investments drive modern generative AI.
  • AlphaFold represents a profound, real scientific triumph in protein folding.
  • Current AI market dynamics mirror historical hype cycles like the .com boom.
  • Keywords: artificial intelligence, llm, hype cycle, gpt, alphafold
  • Core insight: Generative AI is a powerful, sophisticated software tool currently navigating a classic technology hype cycle driven by massive infrastructure investments rather than guaranteed immediate profitability.

Core insights

7
Mental Modelhigh noveltymoderate evidence

AI capability value concentrates in domains with a well-defined, verifiable objective and an enormous but searchable space — AlphaFold reached 90% accuracy on protein folding (10^35 combinations), computed 200M folds, and released them for free, whereas open-ended chat interfaces rely on the same model class but lack a ground-truth scoring function.

Why it matters

It suggests the leverage point for engineers is not model scale but problem framing: choosing tasks where correctness is mechanically checkable produces defensible, reusable artifacts (datasets, folds, benchmarks) instead of demos. This reframes where to spend prototyping budget.

Generalization

Prefer constructing agent tasks with machine-verifiable reward (tests, compilers, formal checks, retrieval ground truth) over tasks evaluated by human-perceived fluency; verifiable domains produce compounding data assets.

Protein folding is one of biology's most complex challenges, with 10^35 possible combinations.
Open source video
AlphaFold achieved 90% accuracy by 2020, computing 200 million protein folds and publishing them for free.
Open source video
Mechanismhigh noveltymoderate evidence

A structural circular-financing pattern underpins frontier model access: a hyperscaler invests capital into a lab and receives cloud consumption commitments back, so compute revenue and investor capital flow through the same counterparties.

Why it matters

If a vendor's AI revenue is partly funded by its own investment capital, published demand signals are a poor proxy for external unit economics. Engineers planning multi-year cost models on today's token pricing should treat price as subsidized and subject to repricing at the trough.

Generalization

When a platform's capital and consumption flow between related parties, treat its pricing and capacity guarantees as unstable inputs to your cost model and design for repricing or migration.

Microsoft invested billions in OpenAI in exchange for Azure cloud commitments (self-dealing model).
Open source video
OpenAI started as a public research lab but shifted to heavy commercial funding and capital needs.
Open source video
Predictionmedium noveltymoderate evidence

Capital expenditure concentration is itself a macro distortion: hyperscaler data center buildout accounted for over half of S&P 500 value increases in 2024, echoing .com-era overbuilding.

Why it matters

Compute supply, GPU availability, and inference pricing are being driven by capital markets rather than by settled demand. That implies procurement terms and capacity can swing sharply if the investment thesis cools, which changes how much infrastructure risk an application team should absorb.

Generalization

Infrastructure-dependent systems inherit the financing cycle of their infrastructure providers; plan capacity and pricing sensitivity against capital-cycle scenarios, not steady-state curves.

Hyperscaler capex figures representing over half of S&P 500 value increases in 2024.
Open source video
Data center infrastructure buildout resembles the overbuilding of the .com boom.
Open source video
Failure Modemedium noveltymoderate evidence

Human pattern-perception bias (pareidolia) is a systemic interface-level failure mode: users attribute sentience and therapeutic intent to LLMs, and the speaker's stated remedy is to treat LLMs as predictive text engines lacking common sense.

Why it matters

Anthropomorphization is not a user error but a predictable property of the audience, so assistant UX, memory features, persona framing, and emotional-support-style responses carry trust and safety risk that must be mitigated in design rather than documented in disclaimers.

Generalization

Systems serving human-facing generative output should treat perceived agency as an emergent hazard: constrain persona claims, avoid memory that implies relationship continuity, and instrument for over-reliance.

Humans are hardwired to perceive intelligence and faces everywhere (pareidolia), making us susceptible to AI hype.
Open source video
Users becoming emotionally attached to chatbots or attributing human sentience to systems like ChatGPT.
Open source video
Tradeoffmedium noveltymoderate evidence

The tooling market is polarizing: coding assistants such as GitHub Copilot deliver genuine productivity gains while many adjacent AI products fail to reach sustainable profit, and numerous small AI coding tools appear and disappear as majors consolidate.

Why it matters

Dependency selection is now a reliability question. Tools that cannot sustain unit economics disappear or get absorbed, taking APIs, pricing, and support with them, so integration boundaries and exit paths matter as much as raw capability.

Generalization

In a consolidating tool market, prefer thin adapters over deep integration points, and treat vendor viability as an architectural constraint alongside latency and cost.

While coding assistants like GitHub Copilot provide genuine productivity boosts, many ancillary AI products fail to deliver sustainable profit.
Open source video
Proliferation and disappearance of numerous small AI coding tools as major players consolidate.
Open source video
Empirical Resultmedium noveltymoderate evidence

The 2012 deep learning breakthrough came from pairing an existing algorithm class (neural nets, backpropagation) with a 14-million-image labeled dataset and available compute — data curation, not a new algorithm, was the unlock.

Why it matters

It argues that for teams without novel research, investment in labeled data and evaluation corpora is the higher-leverage activity than chasing model or architecture changes, and that capability jumps historically track dataset scale.

Generalization

Capability step-changes follow from data and compute availability applied to a known method; budget accordingly for dataset construction and labeling over method novelty.

The ImageNet competition in 2012 used 14 million labeled images to dramatically improve visual recognition.
Open source video
Geoffrey Hinton's work on deep neural nets and backpropagation laid the foundation for modern deep learning.
Open source video
Practicelow noveltyweak evidence

The Gartner Hype Cycle is offered as a planning instrument for deciding when to commit: AI is placed at peak of inflated expectations sliding toward the trough of disillusionment, and the stated purpose of recognizing the cycle is to avoid over-leveraging on temporary manias.

Why it matters

It gives a defensible heuristic for sequencing adoption — pilot now, standardize later — and for timing irreversible commitments such as data-center capacity or platform lock-in during a period of inflated expectations.

Generalization

Treat adoption timing as a resource-allocation decision under uncertainty: keep commitments reversible while a technology is above the trough.

AI has hit the peak of inflated expectations and is sliding toward the trough of disillusionment.
Open source video
Recognizing hype cycles helps investors and developers avoid over-leveraging on temporary market manias.
Open source video

Deep dives

5

Verifiable objectives as the leverage point for agentic systems

Research question

Which agentic software tasks admit mechanically checkable objectives, and do they exhibit steeper capability improvement per unit of data/compute than fluency-scored tasks?

Why

Determines where to spend prototyping budget and whether AlphaFold-like compounding data assets can be reproduced in software workflows.

Protein folding is one of biology's most complex challenges, with 10^35 possible combinations.
Open source video
AlphaFold achieved 90% accuracy by 2020, computing 200 million protein folds and publishing them for free.
Open source video
Source video

Capital-cycle repricing risk in provider token economics

Research question

How much of current frontier model pricing reflects circular investment flows, and how would application unit economics shift under repricing scenarios?

Why

Application teams planning multi-year cost models need to know if token pricing is subsidized and subject to discontinuity.

Microsoft invested billions in OpenAI in exchange for Azure cloud commitments (self-dealing model).
Open source video
Hyperscaler capex figures representing over half of S&P 500 value increases in 2024.
Open source video
Source video

Anthropomorphization controls in assistant design

Research question

Which interaction design controls (persona constraints, memory boundaries, dependency cues) measurably reduce user over-attribution without degrading task success?

Why

Anthropomorphization is a predictable audience property, making trust and safety partly a UX control problem rather than output filtering.

Humans are hardwired to perceive intelligence and faces everywhere (pareidolia), making us susceptible to AI hype.
Open source video
Users becoming emotionally attached to chatbots or attributing human sentience to systems like ChatGPT.
Open source video
Source video

Vendor consolidation and architectural exit paths

Research question

What abstraction boundary minimizes migration cost when an AI tool is deprecated, acquired, or repriced?

Why

In a consolidating tool market, vendor viability is an architectural constraint alongside latency and cost.

While coding assistants like GitHub Copilot provide genuine productivity boosts, many ancillary AI products fail to deliver sustainable profit.
Open source video
Proliferation and disappearance of numerous small AI coding tools as major players consolidate.
Open source video
Source video

Dataset curation leverage vs. method novelty

Research question

For applied tasks, does investment in labeled data and evaluation corpora outperform model/architecture changes at equal cost?

Why

Budget allocation for teams without novel research; tests whether capability step-changes track data scale more than algorithm novelty.

The ImageNet competition in 2012 used 14 million labeled images to dramatically improve visual recognition.
Open source video
Geoffrey Hinton's work on deep neural nets and backpropagation laid the foundation for modern deep learning.
Open source video
Source video

Article ideas

4

The AlphaFold Pattern: Verifiable Objectives Are the Real AI Moat

The leverage in AI engineering is not model scale but problem framing: tasks with mechanically checkable correctness produce compounding data assets and defensible artifacts, while open-ended chat remains hard to evaluate.

Angle

Contrast AlphaFold's 90% accuracy on a 10^35 search space against open-ended assistants without ground truth; derive where teams should aim agentic prototypes.

Source video

Your Token Price Is a Capital Cycle Artifact

Frontier model pricing and capacity are shaped by circular financing and capex concentration, so application cost models should treat today's prices as subsidized and design for repricing/migration.

Angle

Decompose provider pricing into compute, capital recovery, and margin; show why vendor demand signals are poor proxies for external unit economics.

Source video

Anthropomorphization Is a Design Defect, Not a User Error

Because pareidolia makes perceived agency predictable, assistant UX must constrain persona claims, memory continuity, and dependency cues as first-class trust controls, not disclaimers.

Angle

Reframe trust and safety from content filtering to interaction design; propose instrumentation for over-reliance.

Source video

Thin Adapters in a Consolidating AI Tool Market

As niche AI tools disappear and majors consolidate, vendor viability becomes an architectural constraint: prefer swappable adapters over deep integration to keep commitments reversible.

Angle

Use the Copilot productivity gain vs. adjacent product failures as evidence; define adapter boundaries and exit tests.

Source video

Project ideas

4

Verifiable Agent Benchmark Harness

beyond-evals

Agent tasks with mechanically checkable reward (unit tests, type checks, formal verifiers) show a higher rate of improvement per unit of data/compute than tasks scored by human fluency preference.

Proof of concept

Build a paired task suite: coding repair tasks with unit tests vs. open-ended tasks with LLM/human preference scoring; train/finetune or prompt-optimize on both and compare improvement curves.

Measurement

Success@k per dollar/compute; slope of improvement vs. data/compute; delta vs. fluency-scored tasks.

Source video

Token Repricing Stress-Test

new

Current application token economics invert under plausible capital-cycle repricing scenarios because a material share of provider pricing depends on circular financing and capex-driven capacity.

Proof of concept

Build a model that decomposes provider list price into compute, capital recovery, and margin; simulate repricing and capacity quota scenarios for a reference application; test migration to alternative providers.

Measurement

Break-even price increase; margin variance; time/cost to switch providers before and after abstraction.

Source video

Anthropomorphization Guardrail Kit

gatehouse

Constrained persona framing, explicit memory boundaries, and dependency cues reduce user over-attribution and emotional reliance without reducing task success.

Proof of concept

Implement an assistant with configurable persona/memory controls; run A/B tests against an unconstrained baseline on identical tasks.

Measurement

Attribution survey scores; dependency/over-reliance indicators; task success and satisfaction.

Source video

Thin Adapter Migration Testbed

new

Thin, swappable adapters reduce migration time and code churn when an AI tool is deprecated or repriced compared with deep vendor-native integration.

Proof of concept

Implement the same feature twice—direct SDK integration vs. adapter layer—then simulate vendor deprecation/repricing and measure migration effort.

Measurement

Hours to migrate; LOC changed; test pass rate; feature parity.

Source video

Architectural implications

3

Frontier model access is packaged with cloud-commitment financing between investors and providers.

Before

Model pricing and capacity treated as competitive market outputs that reflect underlying compute costs.

After

Pricing and capacity treated as byproducts of capital deployment by related parties, subject to repricing.

Consequence

Cost models should carry an explicit assumption for price and quota discontinuity, with abstraction over model providers to allow migration.

Source video

The AI tooling market is consolidating while sustaining genuine value only in a few proven categories.

Before

Deep integration with the most capable niche AI tool for a given workflow.

After

Thin, swappable adapters in front of a small set of durable vendors.

Consequence

Reduced blast radius when a dependency is acquired, deprecated, or repriced; slight loss of tool-native features.

Source video

Users systematically over-attribute agency to conversational systems.

Before

Trust and safety treated as content-filtering and output correctness problems.

After

Trust treated partly as an interaction-design problem — persona claims, memory continuity, and dependency cues are first-class controls.

Consequence

Assistants need constrained self-description, deliberate memory boundaries, and instrumentation for over-reliance, not just disclaimers.

Source video

Tradeoffs and failure modes

3

Coding assistant productivity vs. unsustainable adjacent AI products

Benefit

Measurable near-term productivity gains from assistants like GitHub Copilot.

Cost or risk

Adjacent and niche AI tools may not reach sustainable profit, leading to product shutdowns, API changes, and lost integration investment.

While coding assistants like GitHub Copilot provide genuine productivity boosts, many ancillary AI products fail to deliver sustainable profit.
Open source video
Source video

Model scaling vs. physical infrastructure constraints

Benefit

Larger models and supercomputing clusters pursued under neural scaling laws.

Cost or risk

Hundreds of billions in data center and GPU spend creates significant power and land demand and distorts broader economic growth metrics.

Hyperscalers are spending hundreds of billions on data centers and GPUs, creating significant power and land demands.
Open source video
Source video

Conversational engagement vs. anthropomorphization risk

Benefit

Natural, emotionally engaging chat interfaces that users adopt readily.

Cost or risk

Users become emotionally attached and attribute sentience, creating trust, safety, and over-reliance hazards when the system is only predictive text.

Users becoming emotionally attached to chatbots or attributing human sentience to systems like ChatGPT.
Open source video
Source video

Open questions

4

Is the hyperscaler data center buildout justified by realized demand, or is it the .com-style overbuilding the speaker compares it to?

Why unresolved

The summary offers a structural analogy and a capex-share statistic but no utilization, payback, or demand-side evidence.

Research direction

Track inference/consumption growth against capex and depreciation schedules for hyperscaler AI infrastructure; look for utilization disclosures and contract backlogs.

Source video

Does the AlphaFold pattern — verifiable objective plus massive structured search space — generalize to agentic software tasks?

Why unresolved

The summary shows one exceptional scientific result but does not indicate whether the same conditions exist for open-ended agent workflows.

Research direction

Prototype agent tasks with mechanically checkable objectives (tests, type checks, formal verifiers) and compare improvement curves against fluency-scored tasks.

Source video

How much of coding-assistant productivity gain is durable versus novelty effect, and how is it measured?

Why unresolved

The summary asserts a genuine productivity boost without citing measurement methodology, task type, or time horizon.

Research direction

Run controlled longitudinal studies measuring cycle time, defect rates, and review load with and without assistants across task categories.

Source video

What happens to application-level token pricing and capacity guarantees when the capital cycle turns?

Why unresolved

The summary identifies circular financing and capex concentration but does not quantify subsidy levels or contract terms.

Research direction

Model cost sensitivity by decomposing provider pricing into compute, capital recovery, and margin, then stress-test application economics under repricing scenarios.

Source video

Key claims

8
factualVerification needed

Microsoft invested billions into OpenAI in exchange for Azure cloud commitments.

Evidence

Microsoft invested billions in OpenAI in exchange for Azure cloud commitments (self-dealing model).

Question

What were the stated investment amounts and the corresponding Azure consumption commitments in Microsoft's disclosed agreements?

Source video
comparativeVerification needed

Hyperscaler capex figures represented over half of S&P 500 value increases in 2024.

Evidence

Hyperscaler capex figures representing over half of S&P 500 value increases in 2024.

Question

What method and denominator produce the claim that hyperscaler capex exceeded half of S&P 500 value increases in 2024?

Source video
factualVerification needed

AlphaFold reached 90% accuracy by 2020, computed 200 million protein folds, and published them for free.

Evidence

AlphaFold achieved 90% accuracy by 2020, computing 200 million protein folds and publishing them for free.

Question

What accuracy metric and benchmark set underlie the 90% figure, and how many structures were released in the public database?

Source video
causalVerification needed

The 2012 ImageNet competition used 14 million labeled images and dramatically improved visual recognition.

Evidence

The ImageNet competition in 2012 used 14 million labeled images to dramatically improve visual recognition.

Question

How much of the 2012 accuracy jump is attributable to dataset scale versus architectural or compute changes?

Source video
causalVerification needed

Neural scaling laws drive the push for larger models and massive supercomputing clusters.

Evidence

Neural scaling laws driving the push for larger models and massive supercomputing clusters.

Question

Do observed scaling curves continue to hold for the workloads now being funded, and at what data and cost limits do returns diminish?

Source video
predictionVerification needed

AI has hit the peak of inflated expectations and is sliding toward the trough of disillusionment.

Evidence

AI has hit the peak of inflated expectations and is sliding toward the trough of disillusionment.

Question

What observable indicators would confirm passage through the trough rather than a temporary plateau?

Source video
factualVerification needed

Numerous small AI coding tools proliferated and disappeared as major players consolidated.

Evidence

Proliferation and disappearance of numerous small AI coding tools as major players consolidate.

Question

What is the observed attrition and acquisition rate among AI coding tools over the relevant period?

Source video
opinionVerification not requested

LLMs are advanced predictive text engines lacking common sense.

Evidence

Maintain professional skepticism and remember that LLMs are advanced predictive text engines lacking common sense.

Source video

Connections

5