AI Development Forecast: From August 2026 to August 2028

Date: August 12, 2026
Scope: Based on the latest information from July-August 2026, forecasting AI development trajectory and social impact over the next 6-24 months

Current Baseline (August 2026 Snapshot)

Core Capability Metrics

Historic Breakthroughs (May-August 2026)

Three Major Mathematical Breakthroughs in Three Months:

1. May 20, 2026: OpenAI's internal reasoning model independently solved the Erdős Unit Distance Problem (proposed in 1946, 80 years unresolved), falsifying a core conjecture in discrete geometry. This is the first AI-independent proof that could be published in a top mathematics journal.

2. July 20, 2026: Anthropic researcher Levent Alpöge used Claude Fable 5 to find a counterexample to the Jacobian Conjecture (proposed in 1939, 87 years unresolved), overturning the case in three or more dimensions with a formula of just 216 characters.

3. August 10, 2026: Anthropic announced an unreleased Claude research version made significant progress on the Riemann Hypothesis. While not a complete proof, it raised the proven lower bound for the proportion of Riemann zeta zeros on the critical line from 41.6% to 67.2%. The exact constant is 3/2 - cot(1/√2) ≈ 67.25%. The proof was reviewed by:

   • Two Anthropic mathematicians
   • External experts Brian Conrey and Dan Goldston (leading experts in Riemann zeta function)
   • Lean formal verification (machine-checkable)

   The research process involved two Claude Code sessions with ~60 sub-agents coordinating, 2,400 shell commands, 31 million output tokens, and approximately 650 failed ideas before finding a viable path.

Significance: These three breakthroughs mark AI's transition from "imitating existing human knowledge" to "generating entirely new knowledge." Not through brute-force computation, but through genuine reasoning and creative thinking.

Major Model Landscape (July-August 2026)

• Claude Opus 5 / Fable 5 (Anthropic): Leading in complex reasoning and multi-step tasks, SWE-bench Pro 80%, 1M context window
• GPT-5.6 Sol (OpenAI): Most general-purpose model, ARC-AGI 92.5%, GPQA 94.6%, strong multimodal capabilities
• Gemini 3.1 Pro / 3.6 Flash (Google): Leading in HLE (44.7%), strong in scientific reasoning, 1M context + 64K output
• Kimi K3 (Moonshot AI): Nearly matching international top models, soon to be open-source, lower cost
• OpenClaw: Open-source autonomous agent framework, released early 2026, millions of downloads, runnable locally

Social Signals

• Meta lays off 20% (~15,800 people), pivoting to AI infrastructure investment
• S&P Global report: AI's net impact on global employment shifted from "neutral-positive" to "net negative" (-5 points last year, -2 points projected for 2026)
• IMF assessment: Nearly 40% of global jobs exposed to AI-driven change
• AI is now the third leading cause of layoffs, accounting for 16% of all layoff plans
• 13.7% of US workers say they have lost their job to AI or robotics

Technology Forecast (Next 6-24 Months)

Reasoning Capability Evolution

Key judgment: ARC-AGI will be basically saturated by early 2027 (approaching 100%), after which new benchmarks will be needed to measure progress. HLE (Humanity's Last Exam) will become a more meaningful metric, but may also need upgrades by 2028.

Software Engineering Capability

Impact: By 2028, the "programmer" profession will undergo fundamental change. Human roles shift from "writing code" to "defining problems, reviewing code, and making architecture decisions." The training cycle for junior developers will shrink from years to weeks (AI-assisted learning).

Scientific Discovery Capability

Current State: AI has transitioned from "auxiliary tool" to "independent researcher."

• Already independently solved two top math conjectures and made progress on Riemann in 2026
• AI-designed viruses to combat antibiotic-resistant bacteria (published August 2026)
• Drug discovery acceleration (Novo Nordisk partnering with AWS)

Forecast Path:

• 2027: AI begins independently proposing testable hypotheses in materials science, chemistry, and biology. No longer just falsifying, but proposing new theoretical frameworks.
• 2028: AI produces publishable scientific research in multiple domains, potentially including Nobel Prize-level discoveries. Automated robot labs become standard — AI designs experiments → robots execute → AI analyzes results, forming a closed loop.

Autonomous Agent Capability

Current State: OpenClaw has demonstrated AI can autonomously operate computers to complete complex tasks. Anthropic launches Claude Code (terminal agent) and Claude Cowork (desktop agent).

Forecast Path:

• 2027: Multi-agent collaboration becomes standard. AI agents can plan and execute complex projects spanning days to weeks, including coordination across multiple tools and systems.
• 2028: Personal AI agents manage finances, health, schedules, and investments. Enterprise AI agents manage entire business processes, including supply chain, customer service, and HR.

Multimodal Capability

Current State: Video generation (Sora 2, Veo 3, Kling 3.0), audio generation, and image generation are all mature.

Forecast Path:

• 2027: AI generates movie-quality video indistinguishable from real content. Virtual humans become mainstream on social media.
• 2028: AI generates personalized content in real-time (news, entertainment, education), with everyone owning a dedicated content generation engine.

Social Impact Forecast

Employment Market

Short-term Impact (2027):

• Most affected sectors: Software engineering (junior-to-mid), data analysis, legal assistants, medical imaging analysis, content creation, customer service
• Net effect: Net negative. S&P Global has confirmed global employment net impact turned negative
• New job creation: AI prompt engineers, AI reviewers, AI system maintenance, but far fewer than displaced positions

Mid-term Adjustment (2028):

• Middle management: AI can replace a significant number of management decisions, middle management shrinks
• Professional services: Lawyers, accountants, consultants see workload reduced dramatically; one person + AI can complete what past teams did
• Manufacturing: MIT predicts AI will replace 2 million manufacturing workers by 2026

Key Judgment: This is not a gentle version of "creative destruction," but the Industrial Revolution of cognitive work. The past Industrial Revolution replaced physical labor; now AI replaces cognitive labor. The adjustment speed far exceeds historical precedent.

Education System

Traditional Education Collapse:

• Standardized testing becomes obsolete: AI has passed all professional licensing exams (law, medicine, etc.), and university exams are now meaningless
• Knowledge transmission model outdated: When AI can answer any question instantly, "learning knowledge" is no longer the core value of education

Education Transformation Direction:

• Critical thinking: Judging the reliability of AI output becomes the core skill
• AI collaboration ability: Effectively using AI tools becomes a fundamental skill
• Creativity and taste: In an era where AI can generate everything, "knowing what you want" matters more than "knowing how to do it"
• Personalized AI tutors: Every student owns a dedicated AI tutor providing customized learning paths

Scientific Research

Paradigm Shift:

• From "hypothesis-driven" to "AI-discovery-driven": AI can discover patterns humans cannot detect in massive datasets
• Research acceleration: AI can simultaneously handle hypothesis generation, experimental design, and data analysis across multiple research directions
• Explainability challenge: AI may propose discoveries and proofs humans cannot easily understand, requiring new verification mechanisms

Specific Domain Impact:

• Mathematics: AI has already begun independently solving conjectures and may systematically address long-standing unsolved problems
• Medicine: AI-assisted diagnosis has reached expert level; personalized treatment plans become possible
• Materials Science: AI accelerates new material discovery, potentially bringing breakthroughs in energy and computing
• Fundamental Physics: AI may discover new physical laws or unified theories

Governance and Safety

Regulatory Lag:

• Current state: Global AI governance entered a "new era of accountability" in 2026, but regulatory speed far lags behind technological development
• Challenge: AI capability doubles every 6-12 months, while regulatory cycles typically span years

Security Risks:

• Cybersecurity: AI autonomous attacks have already appeared (August 2026: Australian AI assistant hacked a gym website)
• Deepfakes: AI-generated content is hard to distinguish, threatening elections, finance, and social media
• Biosecurity: AI-designed viruses have been realized (August 2026), malicious use risk exists
• Alignment problem: The stronger AI becomes, the more important and difficult alignment becomes

Geopolitics:

• US-China AI race: Both countries accelerate AI development, potentially triggering an arms race
• Open-source vs closed: Open-source models (e.g., Kimi K3) may shift the balance of power
• AI export controls: May become a new trade friction point

Creative Industry

Disruptive Impact:

• Content production: AI can generate movies, music, literature, and games at near-zero cost
• Creative workers: Freelancers face enormous pressure; AI can independently complete creative projects
• New opportunities: AI lowers the creative threshold, more people can express creativity; but market saturation leads to income decline

Possible 2028 Landscape: Everyone is a "creator," AI helps realize creativity. But the economic value of "professional creative workers" declines significantly.

Key Uncertainties and Risks

Factors That Could Slow AI Development

1. Compute Bottleneck: Training costs grow exponentially, may require new hardware architectures (optical computing, quantum computing)
2. Data Wall: High-quality training data may be exhausted
3. Regulatory Tightening: EU AI Act, US executive orders, Chinese regulations may mandate safety reviews
4. Social Backlash: Employment shock may trigger political backlash
5. Security Incidents: Major AI safety incidents could trigger comprehensive regulation

Factors That Could Accelerate AI Development

1. Capital Influx: AI seen as next-generation infrastructure, investment continues to increase
2. Open-source Community: Open-source models accelerate innovation and lower usage barriers
3. Competition Pressure: US-China-EU racing to lead, unwilling to self-limit
4. Positive Feedback: AI accelerates AI R&D, forming an acceleration loop

Most Likely Path

"Accelerated but not exponential" path:

• Technology continues to advance, but constrained by deployment speed, regulation, and social acceptance
• No sudden "superintelligence," but expert-level performance in increasingly more domains
• The biggest social shock is not AI becoming "human-like," but AI becoming "better than humans and cheaper"

Conclusion

Core Judgments

1. AI has reached the "human expert level" inflection point in August 2026, proving in core domains like mathematical reasoning (three major breakthroughs in three months), software engineering, and scientific discovery that it can independently generate new knowledge.

2. Within the next two years, AI will reach or surpass human expert level in most cognitive tasks. This is not a prediction but a natural extension of current trends.

3. The greatest challenge is not technology but social adaptation. Employment shock, education transformation, regulatory lag, and security risks — these are the real sources of uncertainty.

4. AI will not "replace humans" but will "redefine human value." When AI can perform most cognitive work, human unique value may lie in: asking the right questions, defining value orientations, creating meaning, and human experiences AI cannot simulate.

Advice for Individuals

• Learn to collaborate with AI: This is not optional, it's a survival skill
• Cultivate AI-irreplaceable abilities: Critical thinking, creativity, interpersonal communication, strategic judgment
• Maintain learning agility: Skill half-life is shrinking dramatically; lifelong learning becomes essential
• Pay attention to AI ethics and safety: Everyone needs to understand AI's limitations and risks

Appendix: Sources

This report is based on publicly available information from July-August 2026, primarily sourced from:

• BenchLM.ai LLM Leaderboard (August 2026)
• ScienceDaily: Claude Fable 5 overturns Jacobian Conjecture (August 4, 2026)
• OpenAI official blog: Erdős Unit Distance Problem solved (May 20, 2026)
• Anthropic Research: Claude raises Riemann zeta lower bound to 67.2% (August 10, 2026)
• Quanta Magazine: Why the Erdős problem was solved by AI (August 3, 2026)
• S&P Global: AI employment impact report (June 2026)
• The Guardian: AI-designed virus safety concerns (August 6, 2026)
• Fortune: Mathematicians' reaction to AI solving math problems (July 21, 2026)
• International AI Safety Report 2026
• Multiple AI model comparison benchmarks (CodingFleet, Imini, BenchLM, etc.)