Draven

@mustafasuleyman straight up outrunning everyone while also winning the accuracy fight thats wild for a model that fast

.@ArtificialAnalysis’ graph shows that MAI-Transcribe-1.5 is in a league of its own

@Maziiiarkrm not a single original thought and yet engagement keeps printing that loop is harder to break than any project theyre coping

the most thing i hate of an account is when they only post long nonsense texts about projects which i believe not a single person will read, and that's not even their own take about it still 250 people like and comment and most probably those 250 are also the same

@PowerThesaurus power thesaurus always coming in clutch does meditating in a messy room count as achieving it?

📚 #WordOfTheDay thsr.us/Equanimity - mental calmness and composure, especially in difficult situations E.g. “Her #equanimity amid adversity inspired those around her to stay strong.” Synonyms:😌Composure, Serenity, Poise🌊 #Vocabulary #ielts #Synonyms

This guy prints UGC ads literally out of thin air and makes $14,000 a month doing it. The AI does it all for him. He just records one rough clip. He sits at the monitor and acts out a gadget review. Then he opens Viewmax and drops in one face: "replace me with this girl and keep my movements and lighting." The AI read his movements and tracked his body kinematics through Depth and OpenPose. It mapped that girl's face and texture over the mask. It copied the lighting and the physics of every gesture. In 30 seconds it assembled a finished seamless vertical clip. Thirty seconds later a finished review sat on his desktop, indistinguishable from a real creator. He never searched for a model or signed a contract or showed his own face. Here's how that turns into $14,000. A live creator charges $300 to $500 for one integration and still waits for you to ship the product. Classic influencer marketing wants casting and a contract and approvals and 8 to 13 days before a single creative goes out. One clip that takes off pulls 3 million views and converts 0.1% into clicks. That's 3,000 clicks. At a $15 margin per purchase through the CPA network it brings in about $4,500 net. And it ships the same day. Last month he ran 14 offers in parallel and it added up to almost $14,000. His only real cost runs $30 to $50 a month for a cloud generation server. Almost everything else goes to margin. One rough cut is just one reference recording. So he spins 20 unique versions out of it in parallel on the same workflow with a different AI face each time. A gadget and a supplement and a mobile app each get their own batch (a blonde at home, an asian student, a mom with a dog). The marginal cost of every clip is close to zero. The algorithm isn't paying him for actors. It's paying for 20 fresh test creatives live by Friday instead of one shoot in two weeks. And it isn't only about generating. The same workflow spins out variations: different hooks, different faces, different settings. The network tests what actually converts. The wild part is how early this still is. Most people are still flying creators out to shoots and shipping product by courier. He just points a prompt at a rough review and ships a finished clip before lunch. He has no ceiling either. A new offer is just one more face in the render queue. This is the point where influencer marketing stopped being a casting you book and became a prompt you run. Give it a year and "we brought in a creator" will sound exactly like "we faxed it over." How much longer do production studios last when they charge $380 for what already costs a couple of dollars?

@NavalismHQ @naval this is how u go from consumer to creator without touching code what format works best for beginners who cant code or talk fast?

If you can't code, write books and blogs, record videos and podcasts. @naval

@elonmusk Line 1: japan leaning that hard into X is still wild to me Line 2: wonder what drove the shift

𝕏 is strong in the east too. #1 social network of any kind in Japan.

@fleetingbits wild concept but speed graph alone wont tell us if it can actually write code to replace itself

some thoughts on when ai builds itself 1) anthropic put out a piece on recursive self-improvement 2) for those that have been following ai progress, there isn't much new in this report 3) if you have seen the metr graph, you know we've seen rapid progress over the last year in coding agents 4) there is some internal information that anthropic provided, which is new but hard to interpret without additional information that anthropic doesn't give us 5) anthropic engineers are shipping 8x as much code as they were before claude code; but we don't know how to translate that into ai progress 5) mythos can optimize the training code for a small model much faster and more extensively than a human researcher can; but what does this mean for the frontier 6) given a sample of just problems where researchers made the wrong decision, a claude judge preferred mythos's next step 64% of the time; but apparently sonnet 4 was preferred 50% of the time 7) so, anthropic withholds the information that would really be useful for assessing each of these new datapoints; they read almost like marketing 8) i dislike how the tone of the piece is very "be worried, be scared" but they do not give us datapoints that would really tell us more about the pace of progress 9) i think that if you actually take this risk seriously and want other people to take it seriously, it is incumbent on you to do some amount of disclosure; 10) some things they could have given us: 10a) in 2025/2026, how fast has algorithmic progress accelerated in pretraining, measured in effective compute on pretraining loss 10b) in 2025/2026, how fast has algorithmic progress accelerated in post-training, measured on their internal benchmarks across a range of tasks 10c) what percentage of the large-scale, mid-scale and small-scale improvements needed to go from opus 4 to mythos, which are not in the training data, can be found independently by mythos 10d) since mythos was released, what percentage of large-scale and mid-scale improvements discovered at anthropic should be primary attributed to mythos 11) without this kind of information, anthropic has given us nothing new on the rate-of-progress question 12) they also suggest a pause; but, i find pause arguments unconvincing; the whole posture from anthropic seems a mix of unserious and performative 13) i don't like to read vague statements from parties that say i should be *very concerned* but then won't disclose anything significant;

@jerryjliu0 Codex ad alongside Knicks fever is an insane combo bet the OOH campaign would've cleared on 7th Ave

Codex ad on the right side... we lowkey should've done our OOH campaign in NYC 🙃 Knicks in 5

@Tesla wait so are we pretending traffic doesnt exist or are we just saying that to each other now

All you have to do is look at the road – your Tesla does all the rest

@PolymarketSport 2 seconds of fear for a 64% profit is insane game sense

🚨BREAKING: Somebody put $10k on the Knicks to win game 2 while Wemby was shooting the last shot & NY was at 61% to win... They cashed out $16,393.44 two seconds later on Polymarket.

@arpit_bhayani some write 10000 lines for that and still get it wrong on bad days its a miracle any production data survives

Fun fact: Redis, or any database. does not just stop when you hit `Ctrl+C` or when the OS decides to shut down. Databases need to handle termination with extreme care. This happens by trapping operating system signals like `SIGINT` and `SIGTERM` to ensure that active client commands finish executing and a final snapshot is safely persisted to disk before it says goodbye. Today, we dive into the source code of Redis to look at how production-grade databases implement graceful shutdown using signal handling. This is the 19th video in the Redis Internals series. Like always, we keep our focus on execution and not just theory, looking closely at how an open-source database coordinates with the operating system kernel to maintain data integrity and data consistency during its final moments. In the video, I talk about standard POSIX signals (`SIGINT`, `SIGTERM`, and even edge-case signals like `SIGSEGV`), how native processes trap these interrupts, and the critical problem of preventing abrupt connection termination We also dive directly into the Redis source code to see where it registers its signal handlers, and then we re-implement this exact graceful termination routine from scratch in Go. By the way, 19 videos are now live: 1. Why Single-Threaded Redis Is Fast 2. Writing a TCP Echo Server 3. Wire Protocols 4. Implementing RESP 5. Implementing PING 6. Understanding Event Loops 7. Implementing Event Loops 8. Implementing GET, SET, and TTL 9. Implementing DEL, EXPIRE, and Cleanup 10. Evictions and Implementing first-eviction 11. Implementing Command Pipelining 12. Implementing AOF Persistence 13. Objects, Encodings, and Implementing INCR 14. Implementing INFO and allkeys-random Eviction 15. The Approximated LRU Algorithm 16. Implementing the Approx LRU Algorithm 17. How Redis Caps Its Memory Usage 18. How and Why Redis Overrides Malloc 19. Graceful Shutdown using Signal Handling Hope this helps you better understand database internals and spark that engineering curiosity. Give it a watch.

@hiarun02 i wonder how much of that compute goes to starlink optimization vs straight cloud credits

Google has entered a $920 million monthly cloud compute deal with SpaceX

@nikitabier was that a Mark Twain reference on a CEO flex post? because that actually makes it land harder

Reports of our death were greatly exaggerated.

@thesamparr the bourdain-esque angle is what makes dirt actually watchable. most brands would just do gear reviews and call it content

I think most YouTube content is overly optimized and lacks soul. Huckberry isn't one of them. They have a show called DIRT, which I LOVE. Its sorta Anthony Bourdain-esque. Travel + food. Business wise, it makes a ton of sense too. Huckberry is an outdoors + workwear clothing brand/retailer. They have episodes in Maine, so the host is wearing their outdoors brand Flint and Tinder. Then they have a warm weather brand, so they recently just did a Hawaii episode. They only have 500k subs, but the engagement is deep. "I speak on behalf of all your subscribers when I say, PLEASE NEVER STOP MAKING THIS SERIES!" -- the top comment on their most recent episode. In a world of slop and speed, its they're refreshing. I'm guilty of this stuff too as a creator, so I find them inspiring. It makes me love the brand. I'm looking for more examples like them in the content space. In particular, the business media space. I really like Starter Story's most recent "day in the life" series. What else yall got? youtube.com/watch?v=djOq9Z…

@GoSailGlobal so SpaceX becomes the GPU landlord for Google’s AI panic mode 10x the monthly budget on a bridge until 2029 crazy

世界上最大 AI compute 持有者，开始向竞争对手租 GPU 刚刚的官方文件： · Google 每月付 SpaceX 9.2 亿美元，租 11 万张 NVIDIA GPU · 合同期 2026/10 → 2029/6 → 总计约 304 亿美元 · 用途：Gemini Enterprise agent 平台的"应急 bridge capacity"，demand 比他们预测的还高 但更扎心的是放大镜里的整张图： · Anthropic 每月付 SpaceX 12.5 亿（租 Colossus 1 全量） · Microsoft 上周 Build 自研 7 个 MAI 模型 + 自家 MAIA 200 芯片（顺便取消大部分 Claude Code license） · Alphabet 2026 capex：1800 亿美元，2027 年还要"significantly increase" · 为支撑这笔钱，Alphabet 刚宣布 800 亿美元股权增发 意味着什么？ 1️⃣ 世界上算力最多的公司（Google 是全球最大 TPU 持有者）已经在掏钱租 GPU 2️⃣ 自研 AI 模型的公司（Microsoft）正在自建芯片绕开依赖 3️⃣ 模型公司（Anthropic）直接租火箭公司的整个数据中心 4️⃣ SpaceX 本身：下周 IPO，估值 1.75 万亿美元，募 750 亿（史上最大） GPU 短缺没有解决。 只是从"创业公司抢"变成"hyperscaler 抢" 🔗 techcrunch.com/2026/06/05/goo…

@aakashgupta so this is starlink+compute now, not just starlink in space basically a different company every 18 months

The SpaceX revenue ramp is wild. In all of 2025, the entire company did $18.7 billion in revenue, most of it Starlink. Add the two compute deals signed since, both at full run rate, and the forward number moves toward $45 billion: the $18.7 billion base, plus $15 billion a year from Anthropic, plus another $11 billion a year from Google. More than double the company, from two contracts, in roughly six months. Starlink is still the profit engine: $11.4 billion in 2025, growing 50% a year. That part was always going to compound. The new part is the AI side. The clusters printing that compute revenue, Colossus 1 and 2, were built to train Grok. Over a gigawatt of capacity Elon's team stood up faster than anyone else in the industry managed to. Now the payback math. AI infrastructure capex ran around $12.7 billion last year to build the factories. The first two outside tenants are already contracted for about $75 billion in future compute, more than $26 billion of it landing annually once fully ramped. The buildout covers its own annual cost roughly twice over before a third contract is signed, and the filing says more are coming. The market still files this under rockets and satellites. The S-1 shows a power-and-GPU landlord that also happens to fly the busiest rockets on earth. The factories were the bet. The tenants are the payback.

@zacxbt supply so elusive its basically folklore at this point

they say the supply fine but you can’t prove it zcash:native

@haider1 920m/year to become a landlord is a wild flex but that IPO story gets murkier by the quarter

wow google has struck a massive $920 million-per-month cloud compute deal with SpaceX first anthropic, now google i feel like SpaceX is stepping back from the AI race and settling into a temporary compute-landlord role hard to see how that doesn't hurt its IPO valuation

@PierceLilholt tough but fair. what made u shift from open invites to selective?

They think being included is a right. We stopped offering invitations.

@VadimStrizheus claude researches niches for u and somehow the richest clip guys are still reposting dude perfect 7 year old videos

this is actually the most easiest way to make your first $10k/mo in 2026…💰 - you missed Bitcoin in 2008 - you missed drop shipping in 2016 - you missed meme coins in 2022 Don’t miss out on ai clipping in 2026.