Gutenberg Lab — Morning Paper

AI Solves Decades-Old Math Problems, Shifting the Profession from Proving to Understanding

Speakers: Lisha Li (a16z infrastructure partner, former mathematician) × Mehtaub Swanee (OpenAI mathematician, former combinatorics researcher) × Mark Selke (OpenAI mathematician, joined after IMO gold medal result summer 2025)

Bottom line: AI reasoning models are now proving results on open problems in sphere packing, coding theory, and group theory that human mathematicians attempted and abandoned, forcing the profession to reorganize around understanding, absorbing, and communicating results rather than proving them.

3-Line Summary

• OpenAI's Astra model improved the 50-year-old sphere packing upper bound from 2^(-0.599d) to the exact asymptotic d/(2πe)^d, proved the existence of non-SOFIC groups in 15 pages, and solved problems humans gave up on after months of effort
• Humans abandon mathematical approaches after hours or days when they seem unlikely to work, but AI executes without psychological cost and lacks "polluted context windows" (memories linking ideas to failed attempts), enabling fresh retries of the same approach
• As proving becomes less of a bottleneck, the implicit functions of understanding, internalizing, explaining, and positioning results within humanity's knowledge become explicit and primary, requiring the field to center "communal understanding" as the core value

3 Key Points

1. GPT-5 solved a literature search in 5 minutes: Mehtaub and friends spent hours trying to determine if an Erdos problem was still open; GPT-5 found the reference in 5 minutes. They subsequently found 10 more cases where the model could resolve open/closed status. Mark notes "getting an idea to work can be a large portion of the battle"—mathematicians often abandon approaches after months, only to learn years later someone else made it work. The model made "the correct bets"

2. 50-year breakthrough in sphere packing: The best upper bound for sphere packing density was an "extremely ugly" 2^(-0.599d) from a 2-page Russian paper in the 1970s. Astra proved the LP bound is exactly d/(2πe)^d + o(1)^d, explaining the numerical conjecture and showing "we now understand this problem in high dimensions very well." Mark had worked on this for 6 months as a grad student with "absolutely zero progress"; the solution was "a few pages of complex analysis"

3. AI proofs are short and elegant: "A year ago, I would have been very surprised to learn that all of these AI proofs are very short and elegant. You're kind of afraid that they're going to generate all these thousand page things you can't understand. But it's been kind of the opposite. Only humans can generate 200 page proofs right now." The non-SOFIC groups proof was 15 pages (the related Aldous-Lyons disproof was 250+200 pages)

Editorial Perspective

When proving becomes "just execution," the scarcity in mathematics shifts to understanding and judgment. This is the classic platform transition pattern—when implementation friction disappears, deciding what to build becomes the rare resource. The mathematical community's recognition that "communal understanding will become an increasingly explicit and valuable part of the subject" mirrors how the rise of the internet made developers shift value from "ability to code" to "knowing what to build." But mathematics has a ceiling: problems like P vs NP may remain unsolvable "even if AI continues getting exponentially better," and the field may reorganize around these enduring mysteries rather than the smaller ones now becoming routine.

Source: a16z Podcast "OpenAI Researchers on the Future of Mathematical Reasoning" (September 8, 2026)

https://a16z.simplecast.com/episodes/openai-researchers-on-the-future-of-mathematical-reasoning-Jn50EkyR

AI Disclosure: Produced with AI assistance; facts and analysis reviewed by our editorial team.

• --

College Kills Heresy — Replit's Founder on the Missing Middle Between Rigid Degrees and Premature Funding

Speakers

Host (a16z Podcast) × Amjad Masad (Founder & CEO, Replit) × Gagan Biyani (Co-founder & CEO, Horowitz and Andreessen Academy)

Bottom line

Young people's civilizational value lies in questioning established assumptions, but modern college trains conformity while VCs hand 18-year-olds a million dollars and demand ten-year commitments — the missing middle is a two-year exploratory environment with no grades, no attendance requirements, and no pressure to optimize prematurely.

Three-Line Summary

• Replit founder Masad argues "the utility of young people is to question deeply held beliefs, to be heretics," criticizing modern college for "beating any outside thought out of them"
• Horowitz and Andreessen Academy's Biyani identifies an "unfair choice" between rigid four-year degrees and premature startup funding, proposing a two-year exploration period with no grades or tests
• Masad warns that "learning the ways of the past" creates fundamental disadvantage as AI enables "self-driving companies" where agents function as organizational glue and bureaucracy becomes invisible machine work

Three Key Points

1. Premature optimization is the root of all evil: Masad compares VCs giving Stanford students "a million dollars pre-idea" to predatory music industry contracts. The right idea "has such gravity, you're not gonna be able to avoid it" — he "tried not to start Replit, actively not to start it" but "it kind of exploded into a startup on its own." The investor track "should be way underutilized. Only a very small number of teenagers should be raising money," with most needing two years to explore without ten-year commitment pressure at 18 [00:17:31]

2. AI enables the "self-driving company": Masad analyzes past experiments like Holacracy and GitHub's no-managers structure as having "failed due to technology limits, not concept flaws." Now "agents can function as the glue" and "have the bureaucracy be this invisible thing happening by machines" — people get true autonomy while AI judges impact. He's already changed how he runs Replit based on this vision [00:24:40]

3. Trust is Silicon Valley's structural advantage: Biyani's local politics volunteering revealed that verbal commitments (50K dollars) in nonprofits have 20-50% follow-through versus 75-100% in Silicon Valley — "by the time you've delivered that check, I've made six weeks of progress." This trust foundation should extend to students: "trusting students to do what they want is something we can uniquely offer" [00:29:47]

Editorial Perspective

Beyond the binary of college critique versus startup worship, this conversation identifies a third path: the "grace period for exploration." Masad's claim that "meandering modern society thinks is useless is actually the most useful" — citing Steve Jobs's calligraphy class that later influenced Mac fonts — defends learning whose utility cannot be measured short-term. Yet his warning that "learning the ways of the past creates fundamental disadvantage" assumes AI will fundamentally reshape how organizations operate. The apparatus reconciling these two — a two-year period with no grades or graduation requirements — reads less as an educational prototype than as securing time for young people to find their own gravity before committing to a trajectory.

Source: a16z Podcast "Amjad Masad on Rethinking College for the AI Era" (September 23, 2026)

https://a16z.simplecast.com/episodes/amjad-masad-on-rethinking-college-for-the-ai-era-sOmuMQGQ

AI Disclosure: Produced with AI assistance; facts and analysis reviewed by our editorial team.

• --

Can plug-in batteries break the <1% residential storage adoption ceiling by eliminating deployment friction?

Featuring

David Roberts × Cole Ashman (Founder & CEO, Pila Energy; former Tesla Powerwall engineer)

Bottom line

The binding constraint on residential battery adoption is deployment friction, not technology—plug-in batteries unlock the blank space of renters and condos, but weak consumer economics ($50-60/year savings) mean the real value lies in utilities and commercial real estate owning them as distributed infrastructure.

3-Line Summary

Whole-home batteries (Tesla Powerwall, etc.) remain at <1% penetration after a decade, with >50% of costs being soft costs (electrician, permits). Pila Energy's 1.6 kWh plug-in battery ($1,500) eliminates electrician install and targets the ~50% of US housing stock (renters, condos) excluded from whole-home battery economics. But at ~$940/kWh (same as Powerwall 3) and $50-60/year bill savings, consumer value centers on backup power, and the company predicts migration to utility/commercial ownership as "default infrastructure."

3 Key Points

1. Deployment friction, not technology, is the bottleneck

Ashman (former Tesla Powerwall engineer) argues that after a decade, <1% of US homes have whole-home batteries because "weeks of back and forth with contractors and construction" and soft costs (>50% of total) are the barrier, not product quality. Pila structurally eliminates this by requiring no electrician.

2. Per-kWh cost equals Powerwall, but market is blank space

At $1,500 for 1.6 kWh (~$940/kWh), Pila costs the same per kWh as installed Powerwall 3. Ashman concedes this openly but argues the product unlocks "blank space—the ~50% of US housing stock (renters, condos) that cannot get whole-home batteries regardless of price." Actual shipped units: 60% homeowners, 40% renters.

3. Consumer economics are weak; real value is infrastructure

With 20¢ peak/off-peak spread and 10W standby draw, annual savings are $50-60. Ashman admits "framing these things as money printers is maybe a bit disingenuous" and that backup power is the primary motivator. But he predicts utilities, commercial real estate, and insurance providers will own/subsidize batteries because "lifetime value is so high that these batteries are worth multiples of the cost to build them."

Editorial Perspective

This embodies the migration of clean energy from consumer product to infrastructure. Ashman's line—"people don't care about kilowatt hours and electrons. They care about hot showers and cold beer"—captures the essence: clean energy should disappear into infrastructure, not be a consumer burden. Eliminating deployment friction is not a technology breakthrough but a speed-of-deployment race, and his claim that "the grid edge is won by whoever kills deployment friction" reflects the platform transition principle: not technology superiority, but deployment speed determines control. However, weak consumer economics mean the product can only succeed if it migrates to utility/commercial ownership models—making this less a consumer play than an infrastructure bet.

Source

Volts, "Do small residential batteries make sense?" (September 23, 2026)

https://www.volts.wtf/p/do-small-residential-batteries-make

AI Disclosure: Produced with AI assistance; facts and analysis reviewed by our editorial team.

• --