Prediction Tracker

Reasoning

Contrarian: AI coding tools may hit a ceiling on what they can reliably automate. Like RPA in 2018-2022, initial excitement gives way to disillusionment as maintenance costs and hallucination-driven bugs accumulate. Consensus says this market grows forever.

Resolution criteria

Resolves YES if the combined revenue growth rate of the top 5 AI coding tool providers (GitHub Copilot, Cursor, Codeium, Tabnine, Amazon CodeWhisperer or successors) falls below 15% YoY as reported in earnings calls or credible market research by end of 2030.

Reasoning

Contrarian: AI makes junior devs far more productive, so companies hire MORE of them at lower cost rather than fewer. The 'AI replaces juniors' narrative ignores Jevons paradox — cheaper labor input means more demand for it, not less. Consensus says junior hiring collapses.

Latest evidence (LR: 0.50)

Oracle's massive 30k job cuts signal a structural hollowing out of traditional engineering roles

Resolution criteria

Resolves YES if aggregate entry-level software engineering headcount at Meta, Apple, Amazon, Netflix, Google (or successors) in 2029 exceeds their 2024 levels, per LinkedIn Workforce Reports, earnings disclosures, or credible third-party analysis.

Reasoning

Contrarian: When everyone can use AI tools, AI fluency stops being a differentiator and the premium collapses — just like 'internet skills' in 2005. Consensus says AI-fluent engineers command massive premiums indefinitely.

Latest evidence (LR: 1.20)

AI fluency is becoming table stakes as 'learning by doing' with AI tools is required to maintain relevance.

Resolution criteria

Resolves YES if Levels.fyi, Glassdoor, or a major compensation survey (n>5000) shows <$10K median total comp difference between 'AI engineer' and 'software engineer' roles at comparable experience levels by end of 2029.

Reasoning

Contrarian: Enterprises care about liability, support, and integration — not model weights. As coding agents handle more critical workflows, enterprises will pay the proprietary premium for contractual guarantees. Open-source wins the hobbyist market but loses the enterprise. Consensus says open-source inevitably wins.

Latest evidence (LR: 0.50)

Fractional node splitting tools make open-weights models like DeepSeek V3 highly accessible, helping developers avoid proprietary lock-in.

Resolution criteria

Resolves YES if enterprise AI coding tool market share (by revenue) for proprietary-model-based tools exceeds their 2026 share, per Gartner, IDC, or equivalent analyst report published by 2031.

Reasoning

Contrarian: AI models trained on vulnerability databases will internalize secure patterns faster than humans learn them. The 'AI code is insecure' narrative is based on 2023-2025 benchmarks; models improve while human error rates stay constant. Consensus says AI makes code less secure.

Latest evidence (LR: 0.80)

Warnings about existential security risks from integrating AI-generated extensions suggest AI code may harbor more vulnerabilities, not fewer

Resolution criteria

Resolves YES if a peer-reviewed study or major security vendor report (Snyk, Veracode, Synopsys) finds AI-generated code has fewer CVEs per KLOC than human-written code in equivalent production codebases by end of 2029.

Reasoning

Contrarian: Every wave of developer productivity tools (compilers, IDEs, cloud, open-source) has increased total developer employment. AI is a productivity multiplier, not a replacement. More software gets built, creating more jobs. Consensus says mass developer unemployment.

Latest evidence (LR: 0.60)

Oracle's mass domestic layoffs paired with aggressive H-1B filings suggest overall engineering headcount growth is being actively suppressed by cost-cutting and automation.

Resolution criteria

Resolves YES if Bureau of Labor Statistics (US), Eurostat, or a credible global estimate (Evans Data, Statista) shows total software developer employment in 2030 exceeding 2025 levels.

Reasoning

Contrarian: Autonomous agents sound great in demos but create unmaintainable code at scale. Technical debt from AI-generated code will force high-profile rollbacks, similar to microservices backlash. Consensus says autonomous agents only expand from here.

Latest evidence (LR: 1.50)

Developer fatigue with AI marketing and reality checks on fully autonomous coding agents increase the probability of rollbacks

Resolution criteria

Resolves YES if 3+ companies with >1000 engineers publicly announce scaling back, pausing, or discontinuing autonomous AI coding agent programs, documented in press releases, blog posts, or credible journalism by end of 2028.

Reasoning

Contrarian: DeepSeek, Alibaba, and ByteDance are investing massively in code models. China has more developers than the US. Export controls push China to build independent tooling that captures the non-US market. Consensus focuses exclusively on US-based tools.

Latest evidence (LR: 1.50)

The democratization of DeepSeek V3 access accelerates Western adoption of Chinese frontier models.

Resolution criteria

Resolves YES if a Chinese-origin AI coding tool (by a China-headquartered company) holds the largest market share by MAU or revenue in any 3+ non-US countries with >100K developers, per credible analyst report by end of 2029.

Reasoning

Contrarian to the contrarians: despite AI hype, natural language is fundamentally ambiguous. Formal languages exist because precision matters. Engineers will still write code — just different code (prompts-as-code, constraint specs, test harnesses). The 'end of programming' narrative is wrong. Consensus says natural language replaces code.

Latest evidence (LR: 1.30)

Interacting with AI is becoming a technical systems engineering task rather than simple natural-language prompting

Resolution criteria

Resolves YES if >50% of production software artifacts at Fortune 500 companies are still primarily specified in formal programming languages (not natural language prompts) as of 2032, per developer surveys or enterprise tooling reports.

Reasoning

Contrarian: Non-engineers using AI to build production systems they do not understand will cause failures in healthcare, finance, or infrastructure. This triggers a backlash — professional licensing for software, mandatory code audits, or AI-code liability laws. Consensus says vibe coding democratizes development harmlessly.

Latest evidence (LR: 1.20)

Rising concerns over existential security risks of AI-generated code in legacy architectures slightly increase the risk of catastrophic failures

Resolution criteria

Resolves YES if any G7 country passes legislation or regulatory rule specifically addressing AI-generated code quality, liability, or mandatory review requirements by end of 2028.

Reasoning

Contrarian: IDE-layer AI tools are a thin wrapper over foundation models. When model providers ship native IDE features (which they already are), standalone AI IDE companies get squeezed. Cursor's $50B valuation is the peak, not the floor. Consensus says Cursor becomes the next big platform.

Latest evidence (LR: 1.50)

Massive compute costs and OpenAI's acquisition of TBPN indicate independent AI wrappers face severe financial viability threats.

Resolution criteria

Resolves YES if Anysphere (Cursor's parent) is acquired by another company, merges, or ceases operations by end of 2028. Resolves NO if it remains independent and operational.

Reasoning

Contrarian: US dominance in AI tooling is not guaranteed. DeepSeek R1 already matches frontier US models at a fraction of the cost. European and Asian developers may prefer local alternatives for sovereignty, cost, and latency reasons. Consensus assumes US tools win globally.

Latest evidence (LR: 2.50)

New node-splitting models remove hardware barriers for Chinese models like DeepSeek V3, explicitly accelerating their adoption among developers.

Resolution criteria

Resolves YES if a global developer survey (n>10000) or credible market analysis shows >30% of professional developers primarily using AI coding tools built by non-US-headquartered companies by end of 2030.

Reasoning

Contrarian: AI doesn't just replace junior work — it codifies senior patterns. Architecture review, system design, and debugging heuristics get embedded in AI tools, democratizing expertise. The shortage never materializes because AI fills the gap. Consensus says senior engineers become the bottleneck.

Latest evidence (LR: 0.70)

Capital inflows to aerospace draining senior engineering talent implies the senior shortage will persist rather than be solved by AI

Resolution criteria

Resolves YES if 'senior software engineer' roles do NOT appear in the top 10 hardest-to-fill tech positions in Hays, Robert Half, or equivalent staffing firm reports for 2029 or 2030.

Reasoning

Contrarian: Scaling laws have limits. Training data for code is finite. The next jump in coding ability may require architectural breakthroughs that don't arrive on schedule. A plateau would deflate the entire 'AI replaces programmers' narrative. Consensus says capabilities only go up.

Latest evidence (LR: 1.40)

Theoretical shifts away from hardware scaling and strict usage limits on Claude Opus suggest brute-force scaling laws may be hitting a computational wall.

Resolution criteria

Resolves YES if any top-3 AI lab's flagship model shows <2% improvement on SWE-bench, HumanEval, or equivalent coding benchmarks across two consecutive major model releases spanning >6 months, before end of 2028.

Reasoning

Contrarian: Tech companies already have great developers. AI coding tools give the biggest marginal improvement to companies with weak engineering orgs — banks building internal tools, manufacturers automating supply chains, retailers customizing e-commerce. The productivity gain accrues to the laggards, not the leaders. Consensus focuses on how AI helps Silicon Valley.

Resolution criteria

Resolves YES if McKinsey, BCG, or equivalent consultancy publishes analysis showing non-tech-sector companies captured more economic value (cost savings + revenue) from AI coding tools than tech-sector companies by 2030.

Reasoning

Copilot has first-mover advantage and deep GitHub integration, but Cursor, Claude Code, and others are rapidly gaining ground. The market is fragmenting — yet GitHub's enterprise lock-in and Microsoft's distribution remain formidable.

Latest evidence (LR: 1.30)

Eroded trust in Azure and the rapid rise of local/open-source alternatives threaten Copilot's market dominance.

Resolution criteria

Resolves YES if any single competitor surpasses GitHub Copilot in monthly active paid users or revenue by end of 2027, per GitHub earnings, competitor disclosures, or credible analyst estimates (Gartner, IDC).

Reasoning

The AI-first workflow is already standard at many tech companies. As AI coding tools improve and enterprises adopt them, the generate-then-review pattern will become dominant for routine code. Enterprise inertia slows adoption but doesn't stop it.

Latest evidence (LR: 2.00)

New tools like Agents Observe show engineers shifting toward dashboards to filter and audit AI-generated multi-agent code

Resolution criteria

Resolves YES if a major developer survey (Stack Overflow, JetBrains, or GitHub) reports >60% of Fortune 500 enterprise respondents say AI generates the initial draft for the majority of their new code, by end of 2029.

Reasoning

Progress on SWE-bench has been rapid — from ~2% in early 2024 to over 70% by early 2026. With agentic scaffolding improvements and better models, the 90% threshold is within reach in the near term.

Latest evidence (LR: 1.50)

Meta's open-sourcing of repoprover advances the automated logical verification necessary for AI models to conquer the final complex edge cases in SWE-bench.

Resolution criteria

Resolves YES if any AI coding system (model + scaffolding) scores >90% on SWE-bench Verified leaderboard with a publicly documented submission by December 31, 2026.

Reasoning

The market is overcrowded with AI coding startups competing on thin differentiation. As foundation model providers bundle coding features, standalone tools face margin compression. This pattern has played out in every enterprise SaaS wave.

Latest evidence (LR: 2.50)

OpenAI's TBPN acquisition and unsustainable compute economics signal an active consolidation and death of thin wrappers.

Resolution criteria

Resolves YES if 5+ AI coding tool startups that raised >$10M in venture funding announce shutdown, acqui-hire, or fire sale (acquisition at <1x total funding) by end of 2027, per Crunchbase or PitchBook data.

Reasoning

Prompt engineering is becoming a universal skill rather than a specialization. As AI tools improve at interpreting intent, explicit prompt crafting becomes less important. The standalone role is already declining in job boards.

Latest evidence (LR: 2.00)

Debugging AI process lifecycles proves AI interaction is a systems engineering task, absorbing the prompt engineering role

Resolution criteria

Resolves YES if 'prompt engineer' (as a distinct job title, not mentioned in requirements) constitutes <1% of software engineering job postings on LinkedIn, Indeed, or Glassdoor by end of 2030.

Reasoning

India already has more GitHub users than the US. AI coding tools are especially valuable for developers who learned English as a second language. Free tiers and lower-cost alternatives will drive massive adoption in India's enormous developer population.

Resolution criteria

Resolves YES if a credible survey or platform data (GitHub, JetBrains, Stack Overflow) shows India has more developers regularly using AI coding tools than the US by end of 2028.

Reasoning

Companies need engineers who can build, evaluate, and maintain AI systems — not just use them. This hybrid role combining ML engineering with software engineering is already proliferating at major tech companies.

Latest evidence (LR: 0.60)

AI unbundling into gig tasks and AI fluency becoming table stakes contradicts 'AI software engineer' remaining a breakout specialized title

Resolution criteria

Resolves YES if 'AI software engineer' or equivalent title appears in the top 5 fastest-growing tech job categories per LinkedIn's annual Jobs on the Rise report, BLS data, or Glassdoor's analysis for any year through 2028.

Reasoning

AI tools commoditize routine coding work. While top engineers see pay increases, the floor drops as more people can do basic development. This mirrors what happened to web design salaries after template builders emerged.

Latest evidence (LR: 1.60)

Oracle replacing domestic workers with cheaper visa labor amid mass layoffs validates strong downward pressure on bottom-quartile US engineering salaries.

Resolution criteria

Resolves YES if BLS Occupational Employment Statistics or a major comp survey shows the 25th percentile of US software developer total compensation (inflation-adjusted to 2025 dollars) is lower in 2030 than in 2025.

Reasoning

As AI handles more routine work, the premium on system design, cross-team coordination, and architectural judgment increases. Companies will pay more for the engineers who can direct AI-augmented teams and make high-stakes technical decisions.

Latest evidence (LR: 1.40)

Massive capital influx into high-performance systems engineering will increase demand and compensation for staff+ talent

Resolution criteria

Resolves YES if Levels.fyi data shows median total compensation for Staff/Principal engineers at FAANG-tier companies in 2030 is >20% higher in real terms (CPI-adjusted) than 2025 levels.

Reasoning

AI tools dramatically lower the skill floor for routine web development, expanding the supply of people who can deliver basic projects. Platform data already shows rate compression in commodity coding categories. But complex work may sustain premium rates.

Latest evidence (LR: 1.40)

Commoditization of standard React/CRUD web development supports a drop in routine freelance rates

Resolution criteria

Resolves YES if the median hourly rate for 'web development' category on Upwork or Toptal falls below $25/hr for the majority of quarters in 2029, per platform data or freelancer surveys.

Reasoning

As AI generates more production code with less human review, the attack surface expands. AI tends to reproduce common vulnerability patterns from training data. The sheer volume of AI-generated code makes a large-scale breach statistically likely.

Latest evidence (LR: 1.60)

The highly targeted Axios supply chain attack demonstrates how malicious actors can use social engineering to trick highly-leveraged AI agents into importing compromised dependencies.

Resolution criteria

Resolves YES if a data breach affecting >10M users is attributed primarily to a vulnerability in AI-generated code, per post-mortem analysis, CVE reports, or investigative journalism by end of 2027.

Reasoning

The security industry always follows the market. As AI-generated code proliferates, security vendors are already adding AI-specific detectors. This is a natural evolution, not a bold prediction — it's already happening at vendors like Snyk and Semgrep.

Latest evidence (LR: 1.50)

Shift from application-level vulnerabilities to dev-environment and supply-chain hardening drives enterprise security needs

Resolution criteria

Resolves YES if >50% of the top 10 SAST/DAST tools by market share (per Gartner Magic Quadrant or Forrester Wave) include AI-generated code detection features in their standard product by end of 2028.

Reasoning

As AI agents gain code execution capabilities and access to production environments, prompt injection becomes a critical attack vector. OWASP has already flagged LLM-specific risks, but coding agent exploitation is an emerging subcategory that may earn top-level recognition.

Latest evidence (LR: 1.80)

Cloudflare intercepting ChatGPT output based on React state reveals critical new vulnerability surface between LLMs and app execution

Resolution criteria

Resolves YES if OWASP Top 10 (web or a new AI-specific list that gains equivalent industry adoption) includes a category specifically addressing prompt injection or manipulation of AI coding agents by end of 2028.

Reasoning

AI is strong at generating unit tests from code, but integration and end-to-end tests require deep domain understanding. Early adopters will achieve high coverage for unit tests while integration testing remains largely human-driven, keeping overall averages below 80%.

Latest evidence (LR: 2.50)

Kitchen Loop framework operationalizes 'Unbeatable Tests', validating completely automated testing

Resolution criteria

Resolves YES if a peer-reviewed study, conference paper (ICSE, FSE), or major company tech blog documents >80% of test cases being auto-generated by AI for new feature development at a company with >500 engineers, by end of 2028.

Reasoning

Fully autonomous deployment requires trust in AI-generated tests, rollback mechanisms, and anomaly detection. While the technology exists in pieces, organizational risk aversion and compliance requirements will slow adoption beyond low-risk changes.

Latest evidence (LR: 0.70)

Growing reality check on deployment timelines for fully autonomous agents lowers the probability of autonomous CD pipelines near-term

Resolution criteria

Resolves YES if credible reporting or company disclosures show >40% of production deployments at 3+ companies with >1000 engineers are fully autonomous (no human in the loop) by end of 2029.

Reasoning

AI code review is improving rapidly and doesn't suffer from reviewer fatigue or social pressure to approve. Multiple startups and incumbents are investing heavily. The controlled study evidence is beginning to accumulate.

Resolution criteria

Resolves YES if a peer-reviewed study or credible industry report shows AI code review catching >50% of bugs that escaped human review in a controlled experiment with >100 PRs by end of 2027.

Reasoning

AI assistants serve as always-available pair programmers. Why schedule synchronous pairing sessions when you can iterate with an AI? But some teams value pair programming for mentorship and knowledge sharing, which AI cannot fully replace.

Resolution criteria

Resolves YES if a developer survey or workplace study (n>500) at AI-adopting companies shows self-reported pair programming frequency dropping >50% compared to pre-AI-adoption levels by end of 2028.

Reasoning

AI tools already enable rapid prototyping. Combining code generation, design tools, and deployment automation, solo developers can ship MVPs dramatically faster. This is already happening for simple apps; the trend will extend to more complex projects.

Latest evidence (LR: 1.25)

Discontinuation of Mac Pro underscores shift to cloud AI orchestration, supporting faster idea-to-MVP cycles unconstrained by local compute.

Resolution criteria

Resolves YES if a credible developer survey or platform data (e.g., Vercel, Replit, or Heroku analytics) shows median time from project start to production deployment is <7 days for solo AI-assisted projects by end of 2028.

Reasoning

The developer role is shifting from writing code to directing AI code generation. As AI handles more implementation, the bottleneck moves to clear specification and quality review. This is the natural evolution of the AI-assisted workflow.

Latest evidence (LR: 2.00)

A viral commit of 12,000 AI-generated posts highlights the explosive growth of AI bloat, shifting developer focus toward review and truncation.

Resolution criteria

Resolves YES if a major developer survey (JetBrains, Stack Overflow, or GitHub) reports >50% of respondents spend more time on requirements/review than direct code writing by end of 2029.

Reasoning

The narrative that AI replaces programmers may deter students from CS programs. However, CS degrees have survived every previous 'coding is dead' scare, and the field's breadth (ML, systems, security) provides resilience beyond just coding.

Resolution criteria

Resolves YES if NCES (National Center for Education Statistics) or CRA Taulbee Survey data shows US CS degree enrollment in 2029 is >15% below the 2025 level.

Reasoning

AI is better at generating high-level web code than low-level systems code. Performance-critical, safety-critical, and hardware-adjacent programming requires deep understanding that AI struggles to replicate. Supply of systems programmers stays tight while demand grows.

Latest evidence (LR: 2.50)

Widespread RAM shortages and GPU gamification demonstrate a cultural shift toward developers aggressively upskilling in hardware and systems programming.

Resolution criteria

Resolves YES if Levels.fyi, Stack Overflow survey, or BLS data shows the ratio of systems programming salaries to web development salaries in 2030 exceeds the 2025 ratio by any margin.

Reasoning

Contrarian: AI tools compensate for gaps in theoretical knowledge by providing on-demand expertise. The practical skills bootcamps teach become more valuable when AI handles the CS fundamentals. But companies may still prefer degrees as a signal. Consensus says the degree gap remains important.

Resolution criteria

Resolves YES if a study (n>200) comparing bootcamp vs CS degree hires at comparable companies finds <10% difference in performance reviews, promotion rates, or productivity metrics in the first year, published by end of 2028.

Reasoning

Contrarian: Current valuations (Cursor at ~$50B) are already stretched. If the AI coding market consolidates, some players fail, and growth slows, the $200B aggregate may be a ceiling rather than a floor. Consensus says this market is just getting started.

Latest evidence (LR: 0.60)

A looming violent market correction and unsustainable unit economics threaten the high valuations of pure-play AI coding companies.

Resolution criteria

Resolves YES if the total post-money valuation (or market cap for public companies) of companies deriving >50% of revenue from AI coding tools exceeds $200B by end of 2027, per PitchBook, Crunchbase, or public market data.

Reasoning

Microsoft's playbook has always been to bundle developer tools into its platform. With Copilot already integrated into VS Code, the natural move is to include basic AI coding features in the standard (free) tier to drive Azure consumption and undercut competitors.

Resolution criteria

Resolves YES if Microsoft offers core AI code completion and chat features in VS Code for free (not requiring a separate Copilot subscription) by end of 2027.

Reasoning

Large enterprises have the engineering talent and the domain-specific codebases to train specialized coding models. However, the organizational incentive to open-source such tools is weak, and most enterprises prefer to buy rather than build developer tools.

Resolution criteria

Resolves YES if a Fortune 100 company outside the tech sector releases an open-source AI coding tool or model that achieves >10K GitHub stars by end of 2029.

Reasoning

Each GPT generation has shown massive jumps on coding benchmarks. GPT-4 already scored high on HumanEval, and the trend line suggests the next frontier model will saturate simpler benchmarks while making strong gains on harder ones.

Resolution criteria

Resolves YES if the next major OpenAI model after GPT-4o (whatever it is named) achieves >95% on HumanEval and >80% on SWE-bench Verified within 6 months of public release.

Reasoning

Autonomous project maintenance requires understanding user intent from issue reports, navigating complex codebases, avoiding regressions, and making design decisions. Current agents can fix isolated bugs but struggle with the holistic judgment that maintenance demands.

Latest evidence (LR: 0.70)

Mounting developer fatigue and reality checks on autonomous coding agents reduce near-term prospects for autonomous maintenance

Resolution criteria

Resolves YES if an AI system autonomously resolves >70% of issues (bug fixes, feature requests, dependency updates) on a public GitHub project with >10K LoC and >100 issues/year, sustained for >6 months, with <10% of its PRs reverted, by end of 2028.

Reasoning

Multimodal input is a natural evolution for coding tools — designers can screenshot UIs, PMs can sketch architectures, and developers can voice-describe changes. But text-based interaction has deep inertia in developer workflows, and multimodal adds latency.

Latest evidence (LR: 2.00)

The release of Pluck directly integrates UI structures into AI coding workflows, validating the shift toward multimodal inputs.

Resolution criteria

Resolves YES if AI coding tools with multimodal input (beyond text/code) capture >25% market share by revenue or MAU per a credible analyst report by end of 2029.

Reasoning

API pricing has already fallen dramatically (GPT-4 pricing dropped >10x in 18 months). Competition between providers, hardware improvements, distillation techniques, and efficiency gains virtually guarantee continued steep price declines.

Latest evidence (LR: 3.00)

The sllm platform drops effective access costs for frontier models to $5/month, signaling a massive reduction in token costs.

Resolution criteria

Resolves YES if the per-token API price for a frontier-quality coding model (comparable to best-in-class 2025 performance) is <10% of the January 2025 price for GPT-4 Turbo or Claude 3.5 Sonnet by end of 2028.

Reasoning

The EU AI Act already establishes risk categories for AI systems. Extending disclosure requirements to AI-generated code in high-risk domains is a natural regulatory evolution. The EU has consistently led on tech regulation.

Resolution criteria

Resolves YES if the EU adopts binding regulation (not just guidance) requiring organizations to disclose or document use of AI-generated code in safety-critical software by end of 2029.

Reasoning

Regulatory burden may drive some startups away, but the EU market is too large to abandon. Most companies will comply rather than relocate. The 20% threshold is extremely high — even GDPR didn't cause that level of flight.

Resolution criteria

Resolves YES if credible reporting or startup surveys show >20% of EU-founded tech startups have moved primary engineering operations outside the EU specifically citing AI code regulations by end of 2030.

Reasoning

India has a massive developer base and growing AI talent, but its AI ecosystem lags the US and China in foundation model development. Government initiatives and the domestic market size provide opportunity, but building a globally competitive tool from India remains challenging.

Resolution criteria

Resolves YES if an AI coding tool developed by an India-headquartered company ranks in the global top 10 by revenue per credible analyst report by end of 2030.

Reasoning

Export controls are tightening, and China is building independent AI infrastructure. But coding tools are software, not hardware — fragmentation requires active blocking of tool access, not just chip restrictions. VPNs and cloud routing may prevent full fragmentation.

Latest evidence (LR: 1.50)

Non-US models matching frontier capabilities and hardware bypasses indicate a fracturing of the global AI ecosystem.

Resolution criteria

Resolves YES if by end of 2029, the top 3 AI coding tools used in China share zero overlap with the top 3 used in the US, and developers cannot easily use tools from the other ecosystem, per usage data or developer surveys.

Reasoning

YC is already seeing founders who use AI agents for nearly all coding. The barrier is not capability but willingness to trust AI-generated code for a production startup. Some founders will embrace this fully as a competitive advantage in speed.

Latest evidence (LR: 0.80)

Reality checks on autonomous coding capabilities make products built entirely by AI agents less likely in the near term

Resolution criteria

Resolves YES if 3+ companies from YC batches in 2026 publicly state or are credibly reported to have built their initial product with <100 lines of human-authored code, the rest being AI-generated.

Reasoning

While AI-generated code does need debugging, it is unlikely to become a standalone job category. Debugging AI code requires the same skills as debugging human code. Companies are more likely to fold this into existing SRE or quality engineering roles than create a new category.

Latest evidence (LR: 1.40)

New tools designed to monitor, filter, and debug multi-agent AI teams point to a new specialization in AI debugging

Resolution criteria

Resolves YES if >10 companies with >5000 employees have a distinct job title or team specifically for diagnosing/fixing AI-generated code issues, per job postings or company org disclosures, by end of 2028.

Reasoning

Current best is ~72% as of March 2026. Rapid progress trajectory but the 80% threshold is a significant jump from 72%.

Latest evidence (LR: 1.30)

Meta's repoprover provides a pathway for AI models to deterministically verify code, accelerating their path to achieving high benchmarks on SWE-bench Verified.

Resolution criteria

Resolves YES if any AI system scores >80% on the SWE-bench Verified leaderboard by April 30, 2026.

Reasoning

Anthropic has been on roughly quarterly release cadence. Claude Opus 4.6 launched early 2026. A new frontier model seems likely within this window.

Latest evidence (LR: 1.50)

The leak of 3,000 files from Anthropic's 'Mythos' project hints at major impending architectural developments preceding a new model launch.

Resolution criteria

Resolves YES if Anthropic publicly launches a model described as a major new release (not a minor update) before May 15, 2026.

Reasoning

GitHub has already moved to metered billing. Competition from Cursor and Claude Code is intensifying. More pricing changes are likely.

Latest evidence (LR: 1.80)

The $65 compute cost for $20 subscriptions points to an urgent need for aggressive price hikes in tools like Copilot.

Resolution criteria

Resolves YES if GitHub announces any change to Copilot pricing, plans, or billing structure before June 1, 2026.

Reasoning

AI coding is the hottest VC category. Multiple startups are reportedly in late-stage fundraising. But >0M for 3+ in a single month is a high bar.

Latest evidence (LR: 0.60)

Drying venture subsidies and a violent market correction make near-term >$10M funding announcements less likely.

Resolution criteria

Resolves YES if 3+ AI coding tool startups announce 0M+ funding rounds in April 2026, per Crunchbase or TechCrunch reports.

Reasoning

The benchmark pace has been accelerating. Multiple labs are working on code models. A >75% HumanEval+ result seems achievable.

Resolution criteria

Resolves YES if an arXiv paper published before April 20 reports >75% on HumanEval+ (extended version) for a single model.