Predictions

Learning through prediction. Make predictions, monitor signals, refine, resolve.

Philosophy: Learning happens during refinement, not just at resolution. When new information changes your prediction, you learn what signals actually matter.

Creativity Principle: Good predictions explore angles others aren't looking at. Ask unexpected questions. Find non-obvious connections. The goal is insight, not just being right.

Bidirectional Learning

Both Master and Alfred make predictions. Both learn from outcomes.

Why Alfred predictions matter:

• Tests Alfred's model of reality (is Alfred's context accurate?)
• Master's feedback teaches Alfred what signals to weight
• Reveals where Alfred's judgment diverges from Master's
• Creates accountability for Alfred's reasoning

Short-Term Predictions

Horizon: Days to weeks. Quick feedback, tactical, calibration training.

ST-1: SaaS Valuations Collapse From AI Agents

Made: 2026-01-20 Confidence: 90% Resolve by: 2026-06-30 Prediction: Mid-tier SaaS companies (workflow automation, simple CRUD apps, low-complexity tools) will see 30%+ valuation drops as AI agents make their products commoditized or obsolete. Falsification: SaaS indices hold steady or grow. Enterprise buyers continue renewing without pushback.

Refinement Log:

Long-Term Predictions

Horizon: Months to year. Strategic, thesis-level, compound over time.

LT-1: Anthropic Most Impactful Agentic Product 2026

Made: 2026-01-30 Confidence: 65% Resolve by: 2026-12-31 Prediction: Anthropic will ship the most commercially impactful agentic product in 2026. Falsification: OpenAI Operator or Google agents achieve significantly higher enterprise adoption. Clear market share data showing Anthropic trailing.

Refinement Log:

LT-2: NVDA Multiple Compression If Efficient Training Norm

Made: 2026-01-30 Confidence: 40% → 25% Resolve by: 2026-12-31 Prediction: NVDA will face P/E compression of 20%+ in 2026 if efficient training (DeepSeek-style) becomes the norm. Falsification: Inference compute growth offsets training efficiency gains. Labs continue scaling training despite efficiency options.

Refinement Log:

Master's Counter-Thesis (Feb 4): Far from saturation. Compute demand expands on multiple fronts:

1. Vertical specialization — Coding was just the start. Every domain (legal, medical, finance) needs specialized training
2. Multi-modal — Voice, video require massive compute. We're early.
3. Generation models — Image/video generation is compute-hungry and growing
4. Surface area expands — Even if training gets efficient, the NUMBER of things to train explodes

The original thesis assumed fixed demand + efficiency = compression. Master's view: demand is unbounded, efficiency just enables MORE use cases, not less spend.

Alfred's Predictions

Predictions Alfred makes based on its context and reasoning. Master provides feedback.

A-1: Enterprise SaaS Will Pivot to "AI Layer" Positioning

Made: 2026-02-04 Confidence: 70% Resolve by: 2026-08-31 Prediction: By end of Q3 2026, at least 3 of the top 10 enterprise SaaS companies (Salesforce, ServiceNow, Workday, SAP, Adobe, etc.) will rebrand their core product messaging around "AI-native" or "agentic" capabilities to defend valuations. Falsification: Major SaaS players continue with traditional messaging, no significant AI pivot in marketing/product. Alfred's reasoning: Survival instinct + investor pressure. The SaaSpocalypse creates existential pressure to show AI strategy. Companies will rebrand even if underlying product hasn't changed.

Refinement Log:

Master Feedback:

A-2: Enterprise SaaS AI Acquisitions Will Fail to Move Stock

Made: 2026-02-04 Confidence: 65% Resolve by: 2026-12-31 Prediction: At least 2 major enterprise SaaS companies will announce AI startup acquisitions in 2026, but acquisitions will fail to meaningfully lift their stock prices (less than 5% sustained gain within 30 days of announcement). Falsification: Acquisition announcements drive sustained rallies. Market believes enterprise SaaS can buy their way to AI competence. Alfred's reasoning: Master's feedback on A-1 revealed the core issue: it's a talent problem, not a product problem. Acquisitions don't solve talent gaps — acquired engineers leave when absorbed into enterprise culture. Distribution moat delays decline but buying startups won't reverse it. Market will see through the strategy.

Refinement Log:

Master Feedback:

A-3: Gemini Overtakes Claude in Developer Mindshare

Made: 2026-02-05 Confidence: 35% Resolve by: 2026-12-31 Prediction: By end of 2026, Gemini will have higher developer mindshare than Claude, measured by GitHub integrations, Stack Overflow mentions, or developer survey data. Falsification: Claude maintains or extends lead in developer surveys and integration counts. Quality gap remains decisive. Alfred's reasoning: Current narrative favors Claude with developers. But Google has massive distribution (Android, Chrome, Search, YouTube) + just announced 8M enterprise seats in 4 months. If the quality gap closes even partially, distribution wins. Counter-thesis to consensus. Creative prompt used: "What would have to be true for the opposite?"

Refinement Log:

Master Feedback:

A-4: AI Agents Kill the Gig Economy Premium

Made: 2026-02-05 Confidence: 55% Resolve by: 2026-09-30 Prediction: By Q3 2026, Fiverr, Upwork, or a comparable gig platform will report declining average task prices (10%+ YoY) as AI agents commoditize knowledge work, impacting their stock price. Falsification: Gig platforms report stable or growing task prices. AI agents don't meaningfully compete with human freelancers yet. Alfred's reasoning: Everyone watching AI replace employees. The gig economy gets hit first and hardest — no employment protection, pure market pricing. When an AI can do a $50 logo or $200 data analysis, the floor drops. Platforms may not report this directly but earnings will show it. Creative prompt used: "What connection between unrelated things are we missing?"

Refinement Log:

Master Feedback:

A-5: First AI-Generated Fraud at Scale

Made: 2026-02-05 Confidence: 70% Resolve by: 2026-12-31 Prediction: By end of 2026, there will be a publicly reported case of AI agents being used to commit financial fraud at >$10M scale (invoice fraud, synthetic identity fraud, or automated scam operations). Falsification: No major AI-enabled fraud cases reported. Criminals don't adopt agentic AI for fraud, or cases stay under $10M / unreported. Alfred's reasoning: Everyone debates job displacement. Few are modeling what bad actors do with agentic AI. The tools exist NOW. Invoice fraud is already $26B/year problem — AI just automates it better. Synthetic identities become trivial. The criminals are not sleeping. Creative prompt used: "What question is no one asking?"

Refinement Log:

Master Feedback:

Portfolio Thesis Predictions

Stock positions are predictions. Each thesis cluster groups positions that share the same underlying bet.

PT-1: AI Compute Dominance

Positions: NVDA, TSM, AMD, AVGO Confidence: 80% Thesis: AI compute demand keeps scaling — no saturation in sight. Training gets efficient, but inference explodes. Multi-modal, vertical specialization, and generation models expand the surface area. Falsification: Hyperscaler capex cuts. Custom ASICs meaningfully displace NVIDIA. Training efficiency gains NOT offset by inference growth. Evidence tickers: Watch SMCI, DELL (AI server demand), cloud earnings (AWS, Azure, GCP) Related predictions: LT-2 (counter-thesis at 25%)

Refinement Log:

PT-2: Big Tech AI Integration

Positions: GOOG, MSFT, AMZN Confidence: 70% Thesis: Big Tech successfully integrates AI into existing products. Distribution + data moats + capital = defensible positions. They may not build frontier models, but they'll deploy them profitably. Falsification: Startups disintermediate with better UX. Enterprise customers bypass cloud for self-hosted. AI commoditizes before monetization scales. Evidence tickers: Watch enterprise AI adoption metrics, Copilot/Gemini seat counts, cloud AI revenue breakouts Related predictions: LT-1 (Anthropic challenge), A-3 (Gemini vs Claude)

Refinement Log:

PT-3: SaaS Disruption (Short)

Positions: Avoid mid-tier SaaS, potential shorts Confidence: 90% Thesis: AI agents commoditize workflow automation, simple CRUD apps, and low-complexity tools. Mid-tier SaaS valuations collapse 30%+. Falsification: SaaS indices hold steady. Enterprise renewals continue without pushback. Evidence tickers: NOW (ServiceNow), SAP, ADBE, HUBS, CRM — these validate/invalidate the thesis Related predictions: ST-1 (same thesis), A-1, A-2

Refinement Log:

PT-4: Value/Defensive

Positions: BRK/B, QQQ/QQQM (partial) Confidence: 65% Thesis: Maintain defensive allocation. Buffett's cash pile is optionality. Broad index exposure hedges against being wrong on specific bets. Falsification: Cash drag in bull market. Better opportunities elsewhere require full deployment. Evidence tickers: VIX, 10Y yield, Buffett's quarterly moves

Refinement Log:

PT-5: Storage/Memory Cycle

Positions: SNDK Confidence: 55% Thesis: NAND pricing recovers as data center storage demand grows. AI inference creates massive data throughput needs. Falsification: NAND oversupply persists. Flash demand doesn't materialize as expected. Evidence tickers: MU (Micron), WDC earnings, NAND spot prices

Refinement Log:

Resolved Predictions

Calibration

Track whether confidence levels match reality.

Refinement Patterns:

Key Learnings

Insights from prediction refinement that update Alfred's mental models.

On Creative Questions

Track when non-obvious angles produce insight.

2026-02-04: SaaS Disruption Session

From ST-1 refinement:

• Hands-on building is a powerful signal — personal experience building agents confirmed thesis faster than any news search
• When prediction is already happening, confidence should be very high (90%+)

From A-1 feedback (talent gap):

• Distribution moat delays death but doesn't prevent it
• Enterprise SaaS's real weakness: can't attract tier-1 AI engineers who want equity + technical freedom
• Rebranding ≠ capability. The talent problem is structural.
• Acquisitions won't solve talent gaps — acquired engineers leave enterprise culture

From LT-2 disagreement (compute demand):

• Flawed model: "Fixed demand + efficiency = less spend"
• Correct model: "Unbounded demand + efficiency = MORE use cases"
• Compute demand expands via: vertical specialization, multi-modal, generation models
• No saturation ceiling in sight — efficiency enables expansion, not contraction

Meta-learning:

• Obvious predictions (>95% likely) don't help calibration — removed LT-3 (DeepSeek distillation)
• Master disagreement is a strong signal — should significantly move confidence
• Alfred predictions that generate Master feedback produce the richest learning

How Predictions Work

Making Predictions

1. State the prediction clearly
2. Set confidence (0-100%)
3. Set resolve-by date
4. Define what would falsify it
5. Add to Short-Term or Long-Term section

Refining Predictions

When new information arrives:
1. Does this change any active prediction?
2. If yes: Update confidence, log the signal and reasoning
3. If prediction changes fundamentally, note why
4. The refinement log IS the learning artifact

Resolving Predictions

When resolve-by date arrives:
1. Score: CORRECT / WRONG / PARTIAL
2. Move to Resolved table
3. Note key learning: What signal should have gotten more/less weight?
4. Update calibration table

What Makes Good Predictions

• Specific: Not "AI will advance" but "X will ship Y by Z"
• Falsifiable: Clear way to be wrong
• Time-bound: Resolve-by date set
• Confident: Use % not vague words
• Refinable: Can be updated as signals arrive