Raoul

I spend 20 hours a week extracting the hidden playbooks from top tech founders so you can master them in 60 seconds.

[Raoul @FreeClaude26]

@HighyieldHarry @tszzl

[Raoul @FreeClaude26]

@DeItaone

[Raoul @FreeClaude26]

LONG FIG - ad agencies and creatives love weave because its user interface enables them to harness foundational models in the exact micro steps that align with how they actually run creative workflows.

@figma

[Anson Lin @ansonlin]

can someone explain to me why I shouldn't buy figma stock right now?

they beat revenue projections every quarter since ipo and the stock is down 85%.

i don't know a single designer that uses claude design.

something doesn't add up.

[Raoul @FreeClaude26]

AI doom and gloom is so disconnected from how the real world is using AI, which is like a more powerful Google search at most.

[Polymarket @Polymarket]

NEW: Cloudflare CEO predicts AI will "destroy small business"

[Raoul @FreeClaude26]

Don’t believe everything you see online kids

[Vivek Sen @Vivek4real_]

STOP SPENDING MONEY ON CLAUDE CODE.

CHIPOTLE'S SUPPORT BOT IS FREE:

[Raoul @FreeClaude26]

@Vivek4real_ Don’t believe everything you see online kids

[Raoul @FreeClaude26]

Perplexity's CEO just said the quiet part out loud: "The model is no longer the product."

@AravSrinivas sat down with @HarryStebbings and dropped the most contrarian AI thesis I've heard all year.

10 takeaways that will reframe how you think about this entire market 👇

1. The only metric that matters is "token value per watt per user." Whoever produces the most valuable output tokens for the least power wins. Everything else is noise.

2. Model builders are not safe. If you sell tokens straight out of the model with no harness, you have no business. The model gets commoditized. The orchestration layer is where value lives.

3. Perplexity changed https://t.co/2EsLmv3orc more than any Google PM ever did. AI Mode copied the font, the citations, the inline bolding, the follow-ups. "Except it's still not as good."

4. Micron will be worth more than Meta in 6 to 12 months. Whatever is the bottleneck commands the price. Right now that is memory.

5. Power is THE bottleneck. Not chips. 40 of every 100 data centers are not getting built because of public resistance. That gap is widening.

6. Agents kill objective transactions. They do not touch subjective ones. Booking a flight gets automated. Buying furniture for your living room stays ad-based forever.

7. Export controls helped America short term. They may be building a monster. Forcing China to vertically integrate hardware could create a far more potent competitor. He puts another DeepSeek moment at 20-30%.

8. The real moat is orchestrating across models, not just tools. You will never find GPT-5 inside Codex or Opus inside Codex. They compete. Perplexity runs both.

9. You will always pay for the frontier. One Jeff Dean over five average engineers, every time. But what counts as frontier keeps changing under you.

10. Jensen Huang wakes up every day, tells himself he sucks, and tells everyone around him they are 30 days from going out of business. $5T company. That is the mentality.

The throughline: nobody is comfortable. Not OpenAI. Not Anthropic. Not even him. "Anyone winning today can lose tomorrow."

PS: His sharpest line was about job framing. The doomer "90% of jobs are gone" narrative is a strategic own-goal. You cannot fearmonger about job loss and then complain you cannot build data centers fast enough. Pick one.

Watch the full thing on @HarryStebbings 20VC. The Micron-over-Meta segment is the clip.

[Raoul @FreeClaude26]

@naval

[Raoul @FreeClaude26]

Sergey Brin sat in a room full of founders and casually admitted Google was "a little bit late" on coding.

Then he dropped 10 truths about AI that most lab leaders won't say out loud 👇

1. Specialized models are dying. One general model now tops math, code, and science. Even Brin didn't predict the convergence - it just happened.

2. Intelligence bleeds sideways. Train it on code, its math improves. Teach it images, it reasons through geometry. Capabilities transfer.

3. "Think step by step" looked like the dumbest trick ever invented. It unlocked a massive leap. The breakthroughs hide inside the obvious.

4. Nobody has mapped the edges - not even the labs. Someone asks the model to do a thing, and it just... works. The frontier is unknown territory to its own builders.

5. Superintelligence won't repeal the laws of math. Impossible problems stay impossible, however smart the machine gets. Smarter ≠ omnipotent.

6. Machines conquered chess and Go. Humans didn't quit - they got better. The tool raises the ceiling; it doesn't retire the player.

7. The real game now: use the tool to build the tool. AI watching its own training runs, generating its own data. The flywheel turned inward.

8. Legacy giants (cars, planes) are quietly experimenting on the side. Boring admin gets automated first. Core engineering is next, not never.

9. World models are the missing piece. To do anything a human can, AI must imagine the physical world before it acts.

10. Confidence isn't checking the temperature monthly. Models ship, leads flip overnight. Play the long game or lose your nerve every week.

PS - he gave away the real moat in a single line: "you start to use the tool to build the tool." Whoever compounds that loop fastest wins the decade.

@sergeybrin @Google @Gemini @demishassabis

[Raoul @FreeClaude26]

The most dangerous founder in 2026 isn't the 22-year-old dropout. It's the 40-year-old who finally has the tools to clone himself 1,000 times.

@bryantchou built Webflow into 1% of the internet over 12 years. Now he's back in @ycombinator with @ployai, and his take on the Lightcone pod flips the "young founder" narrative on its head.

10 key insights below 👇

1. Experience became the moat, not the handicap. Everyone can code now. What can't be commoditised is knowing what to build and how to mould it.

2. You no longer clone yourself once. Garry's framing: AI lets you spin up hundreds of versions of yourself, each with your taste and judgement, all running in parallel.

3. The cave years are dead. Parker Conrad spent 2 years in a Mission basement rebuilding Rippling. That whole arc now compresses into days.

4. Go straight to the gold in the idea maze. A second-time founder skips every wall they've already hit and starts where they left off.

5. Don't sell to engineers. They churn on a whim and chase whoever ships the most tokens. Pick a customer with a real, durable pain point instead.

6. Start with a wedge everyone needs. Rippling began as an offer letter generator. Ploy starts with your homepage. The mundane entry point is the trojan horse.

7. Boil the ocean now that you can. Webflow served 50k designers. Ploy targets tens of millions. AI lets you go horizontal in a way that was impossible before.

8. Taste is the scarce input. Boundless model intelligence is useless without the expertise to steer it toward something world-class.

9. Build the harness. The thin layer that gets a general model to nail a specific outcome (what Anthropic did for Claude Code) is where durable value lives.

10. Run the company AI-native too. Every call transcribed, every proposal auto-drafted, every follow-up scheduled. Abundance of time to actually think.

The thread running through it, courtesy of @garrytan, @sdianahu, @harjtaggar and @snowmaker on @LightconePod: you don't have to be 40. You just have to have taste.

PS the line that stuck: AI doesn't replace the founder's judgement, it removes the headcount that used to stand between judgement and execution. The barbarian who's also a mage finally gets to play both.

[Raoul @FreeClaude26]

There is a MASSIVE gap in the rate of adoption in AI between Silicon Valley and the real world right now.

The extent of usage in most businesses outside of tech is as a thought partner on random tasks via the chat interface.

The start ups that can figure out how to productise bridges for the real world to cross the gap will win.

Here NINE reasons why I believe this gap exists👇

1. AWARENESS: the technology underneath LLMs is complex, fast changing and the value they can provide in workflows takes time and energy to figure out. If you are not drinking water from the fire hydrant on X, you are most likely far behind the frontier. And most CEOs and executives do not have time to do this as they have businesses to run, employees to pay and customers to serve.

2. TRUST: foundational models still hallucinate for no good reason. Every error can cost a business customers, revenue and reputation. Every hallucination erodes trust. Entities that hallucinate can never be trusted as the face of a brand.

3. REPUTATION: CEOs have heard stories of "BOTS" such as the "LOBSTER" going rogue and having a "life of their own" which undermines the trust they can place in agentic capability to be integrated into confidential databases and handle critical workflows.

4. CAPABILITY: beyond chat, the MCP and integration capability of LLMs into tooling is still in its infancy meaning the ability to truly build end to end agentic workflow is still not possible in many business environments.

5. ROI: the first wave of AI adoption was built on FOMO. Boards and CEOs investd millioins of dollars into AI subscriptions and token budgets. BUT, minus FOMO, businesses evaluate financial investment decisions using a simple ROI model - what is the return I am getting from this? Will it make me more revenue or reduce my costs? Now that the hype has started to taper and the token invoices are arriving in the finance department inboxes, questions are being asked: what is the upside on this cost? Products that promise a 10x improvement in experience, revenue growth or cost reduction are much easier to sell than products that ride media hype waves.

6. RELIABILITY: if you have ever tried to build an agent, you would appreciate that model evaluation is a BEAST of a task that requires golden data sets and an understanding of machine learning. The foundational models are EPIC, but need a lot of fine tuning to become 99.9% reliable in specific tasks and settings, with a countless number of edge cases that need to be worked through.

7. RELATIONSHIPS: most SMBs/mid-market/enterprise teams are relationship driven organizations. They are not purely performance optimized units of labour. Employees have careers that span decades at one entity in the real world. A CEO is unlikely to replace an employee who has attended their child's birthday party and has accumulated decades of tribal knowledge that isn't documented anywhere, with an AI agent.

8. AUGMENTATION: the fear mongering from the doomers around job losses from AI has created an incentive to NOT be the CEO that replaces employees with AI to save money. That is counter-humanity. PLUS agent capabilities are not close to being able to replace entire white collar FTEs in the majority of workflows and tasks. The narrative should shift to AUGMENTATION not replacement.

9. SECURITY: businesses outside of the US do not want data going to the US. "Why can't the AI be in our country?" is the question every CEO asks. In-memory processing, zero data retention, on-shore servers and data sovereignty are table stakes to building trust and enabling integration of AI into sensitive customer data sets. AWS Bedrock is setting the standard in this regard.

What do you reckon?

@garrytan
@paulg
@MaxJunestrand
@siliconvalleymm
@thisisgrantlee
@winstonweinberg
@chamath
@DavidSacks
@mcuban
@HarryStebbings

[Raoul @FreeClaude26]

@phibrowser Agree this is a big: the layer between an LLM and a human that enables seamless collaboration.

[Raoul @FreeClaude26]

An investor told @thisisgrantlee his idea was the worst he'd ever heard. "No way you'll ever succeed." 5 years later: 100M users. $2.1B valuation. $100M ARR. Run by 50 people. The idea: @gammaapp.

Here's 10 deep insights from a founder who never quits as dropped on the epic @siliconvalleymm podcast 👇

1. Team before idea.

Don't fall for the idea. Fall for the team. The right complementary co-founders shape the idea, not the other way around. Investors bet on two things only: the team and the dream.

2. There is no perfect timing.

He launched Gamma at peak pandemic, between jobs, with a newborn at home. Every reason to wait. He didn't. Regret is the only deadline that matters.

3. Chase energy, not logic.

Gamma ran two ideas in parallel for 6 months. They killed the one they hit a ceiling, and kept the one they couldn't stop thinking about. Pick the idea that wakes you up at 2am.

4. Build distribution into the product.

Incumbents own distribution. You can't outspend them. So you engineer virality into the product itself from day one. Word of mouth is still Gamma's #1 channel today. Zero paid to start.

5. PMF = organic growth + willingness to pay.

Friends will lie to spare your feelings. Usage won't. If people tell their friends AND put a card down, you've found it. Gamma hit 40M users and $20M ARR off this. The competitor that didn't? Tome: 25M users, $3.5M ARR, shut down.

6. Fundraising is a 2-week sprint.

100+ pitches. Under 50% said yes. He pitched 8pm to 2am for two weeks straight. One founder runs point. Stack warm intros so the nos snowball into yeses. Time-box it or it drifts forever.

7. Don't open the floodgates too early.

Raising money is not permission to mass-hire sales. Reinvest in the core. A third of Gamma's early team were product designers. Most VCs would've called that insane. It built the virality everything else rode on.

8. Hire generalists who spike.

A designer who codes ships end-to-end and skips the handoff. That's how you cross $100M ARR with 50 people. Lean isn't a constraint. It's the moat.

9. Pricing is a living experiment.

Free. $8. $90. Set-and-forget pricing is dead. Align price to value, but build a durable business. Never sell dollars at a discount.

10. Be your company's megaphone.

Build an audience by giving value, not pitching your product. And your first 30 days? Spend all 30 talking to customers. Nothing else moves the needle.

The through line: in 2023 Gamma was stuck at 68 users. He hung on. That belief is what changed the trajectory.

You can build anything now. The only question left is what you should build.

PS: An investor calling your idea the worst he's ever heard isn't a stop sign. It's free market research. Grant internalized the feedback, fixed the distribution gap, and turned the rejection into fuel. The best founders make their loudest critic their unpaid advisor.

[Raoul @FreeClaude26]

The moat you were counting on probably died this quarter.

I went through every founder pitch at @ycombinator Demo Day on @tbpn - robots, drones, MRIs, re-entry vehicles, data center cooling, email agents. Same signal in every vertical: agents are dissolving old moats and paying out raw execution.

Top 10 takeaways:

1. The integration moat is gone. @startupandrew (Tasklet) dynamically generates integrations with AI on demand - even for private internal APIs. Bonus: you don't need a perfect vector DB for search anymore, just an agent willing to run 50 queries in parallel.

2. Sell labor, not hardware. @cybermetheus (Eden Robotics) charges ~$10/hr for robot hours actually worked - because customers already have a mental model for that: a human. Wheels + grippers do 80% of industrial work at a fraction of the cost of legs and dexterous hands.

3. "Cheaper than China" is a real wedge. @hpafrisk (Tenet Industries) is porting consumer-electronics + automotive mass-producibility into defense drones. The constraint isn't the warhead (solved in the '60s) - it's range and GPS-denied navigation.

4. Own the full stack to win on reliability. @rhs (Nine Mothers) builds his own belt-fed shotgun system, match-grade ammo, and fuses acoustic + radar + vision because no single sensor is perfect. $1.6M already delivered to DoD.

5. Microgravity manufacturing is earlier in the supply chain than you think. Payton Case (Dispatch Space) isn't making chips in space - he's growing the crystals (up to 1,000x fewer defects) and surviving Mach 25 re-entry to bring them home.

6 Regulatory precedent = speed. Dr. Ephraim Torres (Adialante) gets to $250 MRI scans (vs $2,500+) by selling scans, not machines - 55% margins off 10–12 year hardware life, with a 510(k) path de-risked by precedent.

7. Power - not chips or land - is the data center bottleneck. Ash (Madron) frames cooling and power as the same coin: save on cooling, get more flops per grid permit. Texas: 150GW requested, ~1.5GW approved.

8. Zero to seven figures in 3 months is the new baseline. Tasklet went ~$333K → $7M run rate. But revenue-maxing is the wrong goal for deep infra/defense - go deep into a few customers instead.

9. Build where the labs won't. The boring-but-valuable layer: sandboxes and tooling that agents want, not just things people want. Hard tech is booming because the money flipped after Claude Code.

10. Clone-proofing is a trap - out-execute instead. @garrytan point: yes it can be vibe-coded, but you didn't, they did, and they built distribution. The value is translating 200-IQ models for the 100-IQ world. Make something people want. Boil the oceans.

PS - the "non-technical founder" is officially extinct. PG says stop calling yourself that. Garry put it best: sit in front of Claude Code and you can fix a bug and land a PR today - wasn't true a year ago. The edge now is taste, agency, and knowing what to build.

[Raoul @FreeClaude26]

@pierreeliottlal Nice! Do you have any advice or frameworks for how to talk to users when you are getting started?

The time they give you is often limited, they are getting little in exchange, how can you maximize the valuable insights per minute spoken...

[Raoul @FreeClaude26]

@hanghuang_ Good tips fellas, congrats!

[Raoul @FreeClaude26]

🚀A 4-year-old company just hit $300M ARR while quietly burning 13 TRILLION tokens a month. That's ~451 MILLION tokens per day, per employee.

Inside @harvey - the $11B legal AI taking on the foundation labs. Breakdown 👇

📊 THE METRICS

~$300M ARR - 3x from $100M last August

~960 employees, 12 offices, ~2,000 customers

Tokens: 1T (Jan) → ~13T (now)

DAU/MAU: 36% → ~52%

Queries/user & hours doubling QoQ

Implied intensity: ~451M tokens/day/employee

Growth was "100% product" - the unlock was switching their entire infra to cloud agents.

⚡️ PRODUCTIVITY / OPERATING RULES

Reinvent the company every 6 months - "if you don't constantly change, you die"

Calendar audit anyone who's breaking: is this tied to your #1 prio?

To greenlight anything, write a full paragraph on WHY. Annoyed to write it → it doesn't matter.

Ship at 51% certainty. Most decisions are two-way doors.

🔮 THE FUTURE OF VERTICAL AI

"Every company is going to sell intelligence."

It's an intelligence-allocation problem: right model, right task, right cost.

Frontier models are expensive + general. Vertical models can match them on specific tasks at ~100x less cost.

The real competition isn't legal tech - it's the foundation labs.

🧪 MODEL EVAL

Most vertical benchmarks are garbage - legal's best was "can it pass the bar?" multiple choice.

Coding is the only saturated end-to-end benchmark. Everything else is wide open.

💸 THE BILLABLE-HOUR PROBLEM (for everyone)
"I just spent a billion dollars on tokens. Where's my ROI?"

Vertical companies win by attributing ROI to every token, per task.

Great breakdown by @MollySOShea / @winstonweinberg.

PS - The quiet tell isn't the $300M. It's DAU/MAU climbing past 50% while output length explodes. When usage deepens as the product gets harder to use, you're not selling software anymore - you're selling intelligence by the token. The ROI-attribution layer is the next moat.