Canva's new AI got so popular this year that it forced the company to cut its growth forecast by a third. The more people used it, the bigger the bill got, because too many of those requests ran on models Canva rented from the big AI labs. Growth guidance fell from 30% to 20%, and its biggest backers marked it down from $42 billion to $34.9 billion. Melanie Perkins, the CEO, told shareholders, "We were relying too heavily on frontier models." This week Canva shipped its way out, a model called Brush. Canva's 160-person research lab built it on top of an open-weights model, then trained it on Canva's own templates, editing data from users who opted in, and grades from the company's staff designers. It edits a deck about the way you would. Writes a little code, nudges the layout, checks how it looks, goes again. Canva says it's 46% cheaper per task than the agent it replaces, and up to 69% faster. At Canva's size, 46% is the whole ballgame. 265 million people use Canva every month. Even a dime per user per month in AI costs would add up to $318 million a year. Cliff Obrecht, the co-founder, said the cost to serve a user went "from cents per month to now many, many more cents per month." So Canva has been swapping rented models for its own, one job at a time. It says its in-house style-transfer model runs 23x cheaper than the frontier version, its image model 30x cheaper and its video model 17x. Cost per AI task is down nearly 90% since April. Brush brings that same play to editing, the part where you fiddle with a design until it finally looks right. Renting the big labs' models is how Canva found out what 265 million people actually wanted, and Brush is how it plans to afford giving it to them.

אם אני הייתי אייזנקוט, בשבוע הבא הייתי מודיע שפניי לממשלת אחדות אמיתית. ובמקביל, הייתי משתלט על הכנסת ומתחיל לחוקק. כל שבוע חוק חדש, שמממש הבטחת בחירות של גוש השינוי מצד אחד, ומשנה את חישובי הרווח וההפסד של נתניהו מצד שני. המצב שנוצר אחרי הבחירות הוא פלונטר פוליטי כמעט מושלם: אף גוש לא יכול להרכיב ממשלה, אבל לאף אחד מהגושים גם אין תמריץ אמיתי להקים ממשלת אחדות. נתניהו אומר לעצמו: זו התוצאה הכי טובה שגוש השינוי הצליח להשיג. בבחירות הבאות הפער אולי יצטמצם, ועם קצת מזל אגיע ל־61. בינתיים אני ראש הממשלה, הכול פנאן. מבחינתי אפשר להישאר ראש ממשלת מעבר עוד שנה. מה יש לי להפסיד? מנהיגי גוש השינוי אומרים לעצמם: אין שום טעם להיכנס לממשלת אחדות עם ביבי. אין לו עניין בשותפות אמיתית, והוא ילעס כל מי שייכנס אליו לממשלה, בדיוק כמו שלעס את גנץ. וחוץ מזה, הבטחנו לבוחרים שלנו ועדת חקירה ממלכתית, ביטול חוקי ההפיכה, שוויון בנטל ועוד רשימה ארוכה של הבטחות שאין סיכוי שנתניהו יסכים לקיים. בתורת המשחקים קוראים לזה שיווי משקל נאש: מצב שבו אף אחד מהשחקנים לא יכול לשפר את מצבו באמצעות שינוי חד־צדדי של האסטרטגיה שלו. התוצאה היא שיווי משקל של אי־שיתוף פעולה, שמוביל לעוד סיבוב בחירות. או בקיצור: יאללה מכות. אז איך יוצאים מהפלונטר הזה? כדי לשנות את התוצאה, צריך לשנות את חישובי הרווח וההפסד שמובילים אליה. במהלך המלחמה בעזה עלתה הצעה שלפיה על כל שבוע שבו חמאס אינו משיב את החטופים, ישראל תשתלט על שטח נוסף ברצועה שלא יוחזר. הרעיון היה פשוט: אם המחיר של אי־הסכמה הולך ועולה, גם התמריץ להגיע להסכמה משתנה. להבדיל אלף אלפי הבדלות, אני מציע לגוש השינוי לאמץ את אותו עיקרון פוליטי: להפוך את המשך הקיפאון מבחירה נוחה לבחירה שיש לה מחיר. לגוש השינוי נכס פוליטי משמעותי שנתניהו אינו מחזיק בו: בהנחה שיש לו רוב יציב בכנסת, הוא יכול לשלוט בסדר היום הפרלמנטרי, בוועדות ובחקיקה, גם בלי להחזיק בממשלה. במקום להמתין שנתניהו יסכים לממשלת אחדות, גוש השינוי יכול להתחיל לממש את הבטחות הבחירות שלו כבר עכשיו. כך הוא משיג שתי מטרות במקביל: הוא מגיע לבחירות הבאות עם קבלות על מה שהצליח להעביר בכנסת, והוא משנה את מאזן הכוחות במשא ומתן על ממשלת אחדות. נתניהו אמנם נשאר ראש ממשלת המעבר, אבל הוא כבר לא היחיד שמחזיק במנופי כוח. איך זה עובד בפועל? כבר בשבוע הבא גוש השינוי צריך לבחור יו״ר כנסת חדש, להעביר לידיו את השליטה בוועדות ולהתחיל לקדם סדרת חוקים שיש עליהם הסכמה רחבה בין מפלגות הגוש. לא חוקים פרסונליים, לא חקיקת נקמה ולא ניסיון לכפות חוקה חדשה על חצי מהעם. חוקים ענייניים שמממשים התחייבויות לבוחרים, ושאפשר לקדם גם מול ממשלה לעומתית. הנה הצעת הגשה לתוכנית חקיקה שבועית: - הקמת ועדת חקירה ממלכתית לאירועי 7 באוקטובר. - ביטול חוקי הרפורמה המשפטית והשבת המצב שקדם לה. - חובת שירות לכול, צבאי או אזרחי, בהתאם לצורכי המדינה. - התניית תקצוב מוסדות חינוך בלימודי ליבה מלאים. - ביטול הקצבות לאברכים שאינם מתגייסים. - שינוי סדרי העדיפויות בהטבות דיור, מעונות ומיסוי לטובת עובדים ומשרתים. - הגבלת הריכוזיות והשליטה של קבוצות גדולות בענפי המזון והצריכה. - מתן אפשרות לרשויות המקומיות להפעיל תחבורה ציבורית בשבת. - חוק למאבק בפרוטקשן ובפשיעה כלכלית - הגנה על תקציבי תוכניות החומש לחברה הערבית לא כל החוקים האלה יעברו בתוך שבוע, ולא את כולם אפשר יהיה ליישם בלי שיתוף פעולה של הממשלה. אבל הכנסת יכולה להתחיל לקדם אותם, להעמיד אותם להצבעה ולייצר מציאות פוליטית חדשה. בכל שבוע שבו אין הסכמה על ממשלת אחדות, גוש השינוי מקדם עוד התחייבות לבוחריו. ובכל שבוע כזה נתניהו צריך להחליט אם הוא מאפשר לחקיקה להתקדם, או נושא במחיר הציבורי של התנגדות לה. במהלך ששת השבועות שבהם ינסה אייזנקוט להרכיב ממשלה, יכולים לקרות שני דברים. או שהמאזן החדש ייצור תנאים לממשלת אחדות אמיתית, עם שותפות פוליטית ואג׳נדה ברורה וחיובית לישראל. או שנלך שוב לבחירות. אבל הפעם, במקום להגיע לבחירות אחרי עוד כמה שבועות של מריבות ומשא ומתן עקר, גוש השינוי יוכל להגיע לבוחריו, וגם לבוחרי נתניהו, עם מסר פשוט: תראו מה הצלחנו לעשות בכמה שבועות, בלי ממשלה. עכשיו דמיינו מה נוכל לעשות בכמה שנים.

Warning: Rated PG-13 for sudden loud noises and intense sequences of agent action. Ep 1: @g_korland, CEO of FalkorDB, on why AI agent memory has to forget. Plus the side project in a living room that became FalkorDB. Full episode in the comments. #AIAgents #AgentMemory #GraphRAG

It might sound surprising, but raising a large sum like $20 million can be more detrimental than securing $2 million for a startup. There are specific pitfalls that emerge with too much early capital. One major issue is overspending before you've even validated product-market fit. This can lead to incredible waste and perverted decision-making, missing critical feedback loops. Another common trap is over-hiring. I've personally witnessed this a thousand times: founders bringing on sales teams when they haven't yet mastered selling their own product. How can you teach what you don't know? This approach dilutes focus and can slow down the essential learning curve required for sustainable growth.

If Intel doesn’t mess up this time, Classiq could do for it what Android did for Google. Intel has repeatedly struggled to turn acquisitions into lasting competitive advantages. calcalistech.com/ctechnews/a…

Prior to 2014, software startups carried some risk — which made them an attractive investment. Execution was complex, scaling was difficult, markets were vast and uncertain; the perfect target for venture dollars. Gradually, that risk evaporated. Playbooks emerged for development, operations, and growth. Markets were understood. Software became safe. So, from 2014 to 2022 the venture market engaged in an elaborate work of fiction. By pouring even more capital into the software category, investors manifested a new kind of risk. Rather than idiosyncratic risk ("will it work"), outcomes became dependent on systematic risk ("will it win"). For a while, this allowed venture firms to keep investing in software startups, and the (now mostly artificial) power law shape of returns seemed to hold. Unfortunately, this coincided with the entry of large, liquidity-insensitive LPs — spawning the megafunds. It turns out, "will it win" is a question more obviously answered by funds of $1B or more, where "will it work" is the domain of smaller boutiques. So, as sustained illiquidity shunted allocation from small firms to the megafunds, this mirage of systematic risk became the dominant game. "Everything is computer"; there are no more mysteries for humans to solve. And make no mistake, venture capital has been decimated. Without the focus on infinitely scalable software, avoiding the toxicity of systematic risk, the opportunity set is far smaller. As the junior sibling, venture capital is also vulnerable to the whims and incentives of the megafunds. TBD how that works out, but so far the relationship seems dominated by insecurity and recklessness. However, idiosyncratic risk remains in the nooks and crevices of the economy. For the few firms and individuals still intersted in backing outlandish ideas, those opportunities will always exist. The returns from this part of the market also seem likely to recover to pre-dotcom levels (median DPI ~2x, rather than ~1x) as it falls towards pre-dotcom levels of allocation.

Lonely and Right Keynes said picking stocks is like a newspaper beauty contest: you don't win by picking the prettiest face, you win by picking the face everyone else picks. Ninety years on, that's still the most honest description of my job which is sad because I think I know beauty. Now the tortoise and the hare. Everyone tells it as "slow and steady wins." That's not the story. The story is that for 90% of the race the tortoise is losing. Badly. In public. The hare is at the cocktail party. The hare's LPs are thrilled. The hare is on the list. Then the tortoise wins at the end, and we tell it to children, and we leave out that the tortoise spent the whole race feeling like a fool. So be honest. Would you rather win 90% of the race and lose at the end, or lose 90% of the race and win? Counted in years of your one life, the hare has the better deal. That is venture capital right now. The young partner in a venture firm would rather be the hare and jump to the next firm than be lonely as the tortoise. The consensus deal is the hare. Expensive, crowded, fun. Everyone at dinner knows the company. The marks go up, the capital piles in, and for years you feel like you're winning, right up to the moment the music stops, which is the only moment that counts. The contrarian deal is the tortoise. It pays only if you're contrarian and right, and you find out which at the end. Until then you're on no list, nobody at dinner has heard of your company, and you look slow. Early and wrong feel identical from the inside. That's the whole difficulty. So the real trade was never consensus versus contrarian. It's finding the thing that's contrarian today and consensus in eighteen months. Buy the tortoise, then watch it grow legs. Cheap entry, then the multiple expands when the crowd arrives, and then you're back at the cocktail party being congratulated on your foresight by the same people who thought you were lonely. Lonely and right, then crowded and rich. That's the business. Finding the ugly duckling that's going to become a swan - that's where the alpha is. What's contrarian now that I think flips? Held loosely: 1. Outcomes, not tokens. Everyone is selling the enterprise intelligence by the pound. The enterprise doesn't want intelligence. It wants the claim processed, the invoice reconciled, the ticket closed. One in five companies has AI in production; one in fourteen has rebuilt the factory. Whoever sells the other thirteen a finished outcome, priced like a result and not like a utility bill, wins the decade. It looks like a services business on the way in. So did Amazon. 2. Security that gets paid for the breach that didn't happen. The attackers have agents now, so every company on earth is about to be robbed at machine speed, and the defender's product is still a dashboard. The bet is the company that stops selling the dashboard and starts selling the result: we handle it, we're on the hook. Security and insurance were always the same business. Someone is about to notice. 3. The specialist labs. Not the frontier, which is consensus at a trillion, but the labs that do one thing for one industry at a hundredth the price. 4. The boring half of the physical world. Robotics is loud. Inspection, maintenance and warehouses are quiet, and the revenue lives in the quiet part. I may be wrong about all four. That's rather the point. If I were sure, so would everyone, and then it wouldn't be a trade. It'd be a cocktail party.

Rejection is a fundamental part of success. Anyone that gives up at the first obstacle will likely never bring anything into existence. Catch-22: 22 rejections Dune: 23 rejections The Help: 60 rejections Harry Potter: 12 rejections

I've learned over my two decades in venture capital that money is not just a commodity. The idea that all investors are fungible doesn't align with my experience. VCs hold significant power to influence a company's path. They can either help propel a founder forward or seriously hinder their progress. This understanding drives my decision to get involved very early, well before product-market fit. I want to partner with founders at the outset. My focus is on helping them navigate early experiments, recruit great people, and manage the intense stress that comes with building a startup while trying to live healthily. Follow for more insights on venture partnerships.

This, by @nfx, is truly fantastic. Ep. 3 - 22 Mindsets That Kill Your Speed (NFX Masterclass) buff.ly/M88XOZd

This can't be good. "But pressure to back promising new AI startups is upending that process, with investors sometimes paying wildly different prices for the same company. That means a company can notch multiple valuations simultaneously or in rapid succession, and sometimes only the higher number is made public. These types of deals have proliferated in Silicon Valley, the center of the AI boom, and are now spreading to Europe. That is alarming some founders and investors who say inflated valuations can leave startups with little room to grow into their price tags, particularly in a region with fewer blockbuster exits." buff.ly/JABwaRh

JFrog’s recovery defies the SaaS sell-off as AI makes its software more essential. The Israeli company lost $700 million in market value in minutes after Anthropic’s Mythos announcement, but its shares have since surged. calcalistech.com/ctechnews/a…

REPORTED FIVE-YEAR CREDIT default swap spreads, a measure of what it costs to insure against a borrower's default, are about 101 basis points for Nvidia, the chipmaker, and about 1,090 for CoreWeave, which rents out computing power built on Nvidia's chips, nearly eleven times as wide. A basis point is one hundredth of a percentage point. These are annualized spread quotations, which are not necessarily the cash payments the contracts require. Both quotes come from a corner of the credit market that has grown far busier in the past three months, and the sharp difference between them offers another view of financing risk alongside the increase in trading. buff.ly/cz6MbWp

Germany has found that its AI ambitions are butting up against its climate commitments. EU member states have committed to cutting greenhouse gas emissions by 55 percent by the end of the decade. But policies meant to help meet that target are interfering with attempts to trigger a data center building frenzy. buff.ly/n3dh1HG

MIT engineers’ design for a robot that both swims and flies could lead to a new class of vehicles for ocean exploration. trib.al/IpukNGA

Foursquare was acquired last month. The outcome was disappointing. It's taken me a while to write this, but I think I'm finally done grieving it. Here are some thoughts on our final years, an effort to honor the team who helped build & maintain FSQ's legacy, and my attempt to finally get some closure. teendrama.substack.com/p/som…

New on the blog: — Room for ambition A late-night conversation at a startup office in Finland has stayed with me for almost twenty years. Some thoughts on ambition, getting good at something, and the people around you. karrisaarinen.com/posts/room…

Equivocating ambition w/ massive funding rounds is the cancer of an entire generation of founders and a magnet for scam artists.

Wrote this a few months ago and I send it quite frequently open.substack.com/pub/nikunj…

I've watched a few really good founders raise big seed rounds this year, then struggle to raise their A, even as the business kept making progress. The seed valuation had set a bar they weren't ready to clear yet. And with everyone excited about the term sheet, it was easy to put off what they'd need to show at the next round. One of them had about as strong a resume as you'll see in AI, and a team to match. Sub-$1M ARR would have been a fine A story at a normal seed price. At theirs it wasn't. (For what it's worth, I write both seed and A checks, so I sit on both sides of this.) The top of the seed market is a bit wild right now. Carta's 95th percentile seed valuation was about $72M in Q2 last year. This Q2 it was $200M. Meanwhile the share of seed companies that get to a Series A within two years has roughly halved since 2018. During a raise almost all the attention goes to how much and at what valuation. The question I wish more founders asked is whether the company is actually set up to hit the growth the next round will expect. If you're planning on a 2-3x step-up at the A, here's what that means. A $25M post means pitching a $50-75M A, which a company with real traction can get to. A $100M post means $200-300M, and at that price the A partner is comparing you to the handful of companies growing faster than almost anyone. The bar right now is brutal. In the A conversations I'm in, a $15M seed (usually a $75-100M post today) means $2M+ ARR, growing 5-6x over the last year, with a believable path to another 5x. None of this means raise less. Before you sign, work backward: what will we need to show at the A, and does this money and ~24 months get us there? If it does and the terms are good, take it. If the honest answer is "only if everything goes right", that's worth knowing now rather than in month 20. A big seed buys you runway and a higher bar. Most founders only count the runway.