Ablative Software

6 hours ago

Ablative software is built to fill gaps in current AI models and is consumed when the next generation closes those gaps, as seen with Cognition's price drop from $500 to $20.
Companies like Jasper, Chegg, Pinecone, and Adept illustrate the pattern: their value evaporated when models improved, showing that software built on model shortcomings is a consumable, not an asset.
Frontier labs use the ecosystem of tools built on their models to identify gaps, then close them in the next training run, making today's revenue a measure of future casualties.
The real value lies not in scaffolding code but in non-trainable edges: proprietary data, distribution, liability acceptance, and domain judgment.
Surviving companies treat their software as fuel, not a fortress: they cut prices, rebuild around new models, and charge for outcomes rather than machinery.
Founders should run self-ablation experiments, design for planned obsolescence, and keep an honest ledger acknowledging that model progress is the main driver of software decay.

Startups older than two years likely have obsolete business plans due to rapid changes in AI and markets.
Venture capital has shifted heavily toward AI, making it harder for non-AI startups to raise funds.
AI development tools enable building MVPs in days, reducing the need for large engineering teams and challenging old assumptions about cost and speed.

The most effective software adoption strategy is creating 'building blocks' that enable others to build quantity over quality, exemplified by Ghostty's libghostty reaching millions of users in two months vs. the app's one million daily update checks in 18 months.
Modern software development increasingly involves assembling high-quality, documented components using AI agents, which excel at gluing proven parts together but still require human judgment for quality control.
The building block economy lowers quality barriers, increases awareness via niche use, reduces maintenance burden, and outsources R&D, leading to more stable and purposeful mainline applications.

AI-driven development is breaking traditional code review: faster code generation caps velocity at slow human reviews.
Asynch diff-centered code review, ~20 years old, fails in an era where writing code is cheap but reviewing isn't.
Diffs lack context, requiring reviewers to reconstruct intent and design, adding expensive overhead.

Data scaling is far behind compute scaling because generating high-quality data remains a difficult, bespoke task.
Anthropic's edge comes from focusing on coding models, which allows it to collect vast amounts of labeled data from users who pay for the privilege.
Programming is a key domain, but achieving recursive self-improvement (RSI) requires more than just code data.

Software design is best learned through hands-on experience, not just formal courses.
Conway's Law explains that software architecture mirrors the social structure of the producing organization.
Incentive structures (e.g., PhD deadlines) heavily influence the quality and nature of scientific software.

Startups older than two years likely have obsolete business plans due to rapid changes in AI and markets.
Venture capital has shifted heavily toward AI, making it harder for non-AI startups to raise funds.
AI development tools enable building MVPs in days, reducing the need for large engineering teams and challenging old assumptions about cost and speed.

The most effective software adoption strategy is creating 'building blocks' that enable others to build quantity over quality, exemplified by Ghostty's libghostty reaching millions of users in two months vs. the app's one million daily update checks in 18 months.
Modern software development increasingly involves assembling high-quality, documented components using AI agents, which excel at gluing proven parts together but still require human judgment for quality control.
The building block economy lowers quality barriers, increases awareness via niche use, reduces maintenance burden, and outsources R&D, leading to more stable and purposeful mainline applications.

AI-driven development is breaking traditional code review: faster code generation caps velocity at slow human reviews.
Asynch diff-centered code review, ~20 years old, fails in an era where writing code is cheap but reviewing isn't.
Diffs lack context, requiring reviewers to reconstruct intent and design, adding expensive overhead.

Data scaling is far behind compute scaling because generating high-quality data remains a difficult, bespoke task.
Anthropic's edge comes from focusing on coding models, which allows it to collect vast amounts of labeled data from users who pay for the privilege.
Programming is a key domain, but achieving recursive self-improvement (RSI) requires more than just code data.

Software design is best learned through hands-on experience, not just formal courses.
Conway's Law explains that software architecture mirrors the social structure of the producing organization.
Incentive structures (e.g., PhD deadlines) heavily influence the quality and nature of scientific software.

Hasty Briefsbeta