Tarek retweeted
Do not use your energy to worry. Use your energy to believe, to create, to learn, to think and to grow.
—Professor Richard Feynman
Tarek retweeted
Blackstone CEO Stephen Schwarzman reveals how a 2015 bus ride with Jack Ma led to his firm being the world’s largest data center owner
“We did see the whole AI revolution, and that's just because of an accident. When I was in China in 2015, sitting on a bus with Jack Ma, who is the founder of Alibaba, and he started talking about that.”
“I literally had no idea what he was talking about, and he explained that this would be the most consequential technological change which would change the world.”
“So I have a bunch of academic relationships, one of whom was the head of MIT. He said, ‘Jack's underselling this’. So I said, ‘how do you know?’ He said, ‘because we invented AI at MIT.’”
“And we positioned the firm at Blackstone in this unique way where we bought a very large data center, a company, in 2021 for $10 billion. It's now 15 times larger.”
“We're the largest owner of data centers in the world. That puts us in touch with the six hyperscalers, because we work for all of them. So we know what's happening almost everywhere in this ecosystem.”
Tarek retweeted
من ساعة ما Jev نزل، التايملاين كله بيتكلم عن Decision Models.
أغلب الكلام عن الـ API والـ benchmarks.
إحنا حطينا واحد في أصعب مكان ممكن: جوّه الـ hot path بتاع مكالمة صوتية لايف، والـ budget كله 400ms.
والهدف ما كانش نصلّح حاجة بايظة. كان ناخد حاجة شغالة كويس للمستوى اللي بعده.
الـ Knowledge Base retrieval في Agent.sa مبني على الـ industry default: BM25 + embeddings، وRRF، وfreshness، من غير أي model call في الـ hot path. سريع، وبيجيب الإجابة الصح أول نتيجة في 89.3% من الحالات.
رقم فرق كتير بتقف عنده.
بس في المكالمة الصوتية مفيش صفحة نتايج المتصل يختار منها. فيه إجابة واحدة بتتقال.
فقررنا نروح ورا الـ last mile.
والـ last mile ده ليه سقف معروف في أي retrieval مبني على similarity:
• بيقيس "إيه أقرب entry؟" مش "إيه الـ entry اللي فيه الإجابة؟". "باقة Gold فيها إيه" و"باقة Gold بكام" تقريبًا نفس الكلام، وواحدة بس فيها السعر.
• الـ RRF بيرتّب بالـ rank مش بالـ score، فالـ top 10 بيبقوا جوّه 13% من بعض، وأي tie-breaker زي الـ freshness بياخد وزن أكبر من حجمه.
• مفيش "مش عارف" calibrated. أحسن heuristic ("لو أعلى score واطي يبقى مفيش إجابة") بيدّي AUROC 0.74.
الحل: سيب الـ retrieval يعمل اللي هو شاطر فيه، وحط قاضي فوقه.
الـ retrieval بيطلّع أحسن 8 candidates، والقاضي بيقرا كل واحد ويحكم.
والقاضي عندنا Decision Model مفتوح المصدر: Clef من Cloudflare. Apache 2.0، ومتوافق مع الـ API بتاع Jev.
ولأن الـ weights مفتوحة، شغال على الـ GPUs بتاعتنا، والـ KB بتاع العميل ما بيروحش لأي API برّه.
السؤال اللي بنسأله Score بأربع درجات:
0 ← موضوع تاني
1 ← نفس الموضوع، بس مش بيجاوب (السعر، الشرط، الميعاد)
2 ← بيجاوب جزئيًا
3 ← بيجاوب مباشرةً
الدرجة 1 هي كل الحكاية. هي اللي بتدّي اسم للـ near-miss: الحي صح، والبيت غلط.
ونصيحة: اكتب كل درجة بجملة كاملة، مش رقم. الأحكام بتبقى أحدّ بوضوح.
بعدها الـ reorder دالة deterministic مملة عن قصد:
2 و3 الأول، بعدهم 1 واللي ما اتحكمش عليه بترتيب البحث، و0 في الآخر. Demote، مش remove.
والـ freshness بقت tie-breaker بس.
وحقل answerable بقى tri-state: true / false / null. والـ false هي اللي بتخلّي الـ Agent يقول "خليني أتأكد وأرجعلك" بدل ما يقول أقرب حاجة وخلاص.
قواعد الـ hot path اللي ما بتتكسرش:
• deadline واحد 400ms للـ 8 أحكام مع بعض. اللي يتأخر يتشال، من غير retries. ولو مفيش حاجة رجعت، fallback على ترتيب البحث العادي. قاضي بطيء عمره ما يوقّف مكالمة.
• allowlist مش denylist: 4 حقول بس بتروح للقاضي، وأي حقل يتضاف بعدين بيفضل private لحد ما حد يقرر غير كده.
• off / shadow / live لكل Agent. الـ shadow بيحكم بعد ما الرد يتبعت ويسجّل كان هيغيّر إيه: latency ودقة production بصفر مخاطرة.
• status لكل query: disabled / shadow / applied / partial / timeout / error. القاضي المتعطّل لازم يبان، مش يسكت.
الـ eval، لأننا ما بنشحنش على vibes:
• أسئلة شبه المتصلين: عامية، فصحى، إنجليزي، خليط، تفريغ صوتي مكسّر، وأسئلة الـ KB فعلًا ملهاش إجابة.
• الـ labels عملها موديل تاني، عشان القاضي ما يصحّحش امتحانه بنفسه.
• Ablation: نفس الـ pipeline من غير قاضي طلع مطابق للـ baseline، يعني المكسب كله من القاضي.
• Paired + McNemar: احتمال إن الفرق صدفة حوالي واحد في المليون.
النتيجة:
• Top-1 من 89.3% لـ 97.5%، ومقابل كل سؤال اتأثر سلبًا، 11 اتحسّنوا
• الإنجليزي على محتوى عربي: من 81.7% لـ 98.3%
• السؤال عن تفصيلة محددة: من 83.3% لـ 95%
• "مش عارف": AUROC من 0.74 لـ 0.99
ولأن الـ probabilities calibrated، الـ threshold بقى قرار business مش تخمين.
عند 0.5 بيمسك 96.4% من الأسئلة اللي ملهاش إجابة، وبيغلط في 3.3% بس.
ولو "مش عارف" الغلط عندك أغلى، 0.2 بتنزّل الغلط لـ 1.9%.
وأحلى جزء: "مش عارف" بقت backlog للـ KB.
كل سؤال القاضي قال إنه ملهوش إجابة بيروح لـ worker في الخلفية: يخفي الأرقام الشخصية، يعيد الحكم على اللي فاته الـ deadline، يصنّف (برّه الموضوع، محتاج live data، مش واضح، ولا knowledge ناقصة فعلًا)، يدوّر أعمق من الـ top 8، ويجمّع الأسئلة اللي بتطلب نفس الحاجة.
وهنا الـ trap اللي محدش بيقولك عليه:
نفس السؤال بالعربي والإنجليزي، الـ similarity بينهم من 0.22 لـ 0.43.
سؤالين مختلفين في نفس الموضوع: 0.72.
يعني أي similarity threshold هيشتغل بالعكس بالظبط.
فالـ embeddings بتعمل shortlist بس، والـ Decision Model هو اللي يقرر "ده نفس السؤال؟". وفرّق صح بين رقم واتساب Velents ورقم واتساب Velents Pay.
والاقتراح النهائي للـ KB ممنوع يألّف أي fact. ومش بالـ prompt، بالكود. لأن KB بتألّف أسعارها أسوأ من KB فيها فجوات.
وكل ده على thread مخصص بـ bounded queue، مش background tasks بتاعة الـ framework، لأنها بتشارك الـ pool مع الـ retrieval، ومحدش عايز gap analysis يأخّر مكالمة حد.
الخلاصة:
الـ Retrieval شاطر يلاقي candidates، والـ Decision Model شاطر يختار بين near-misses. دول شغلانتين مختلفتين.
التوليد بقى سلعة.
القرار جوّه الـ hot path هو الهندسة الحقيقية.
Tarek retweeted
If you wonder why everyone fights over pro rata...
Garry Tan's first fund initialized Fund 1 returned 55x.
DPI on the first 98 companies added up to about 1x.
Instacart was another 1x, but they weren't able to take any of their pro rata.
Coinbase, where he kept taking his pro rata through multiple rounds, did the final 53x.
Tarek retweeted
Of course, that’s your contention. You’re a first-year machine-learning engineer. You just got finished readin' Attention Is All You Need, probably watched a Karpathy video too. So now you’re convinced everything is just transformers and scaling laws.
That's gonna last until next month when you discover convolution, and then you're gonna be talkin' about inductive biases and locality and how CNNs were actually incredibly compute-efficient for vision.
Then you're gonna read the FlashAttention paper, and suddenly everything's about IO complexity and SRAM and how FLOPs don't matter because the whole goddamn thing is memory-bandwidth bound.
That'll last until somebody shows you an MoE model, and then you're gonna be in here regurgitating DeepSeek, talkin' about expert parallelism and active parameters and how dense models are economically obsolete.
Then six months from now you'll write one shitty Triton kernel, look at an Nsight trace for the first time, and start telling everybody Python isn't actually the bottleneck because the GPU is asynchronous.
And by next year you're gonna be standing right here explaining to me that your 400-billion-parameter model is fast because you turned on CUDA graphs.
Ben Affleck reveals he writes Python, understands convolutional neural networks, worked extensively with GPUs, and used his celebrity status to get private looks at Google and OpenAI’s video models
“I’ve always been kind of into computers since I was young. Then, when film started to move from analog film to digital, I became more interested in that aspect of it. The visual-effects workflow for many years has included machine learning, so I can write pretty shitty Python scripts and stuff like that.
“With convolutional neural networks, which were the precursors to what the transformer can do, which is much more computation simultaneously, you would do things like look at what’s called a tensor. That’s the numerical translation of a visual image in numbers, like the batch number, the frame number and the red, green and blue values of each pixel in each frame. It’s just that simple. That numeric is called a tensor.
“You’d use a convolutional neural network to identify patterns that reveal what’s called edge detection or feature extraction, which is identifying patterns well enough to know, this is where the window ledge is, so we can more easily take the green-screen image out and replace it with something.
“That was familiar to me early on because, prior to Artists Equity, I had a small visual-effects company. I’ve worked with GPUs a lot too. The visual-effects guys said, ‘Hey, you should see. There are a couple: Google and this other company, OpenAI, are doing really interesting stuff with transformers in video.’
“I’ve learned that I can actually just call up and go, ‘Hey, it’s Ben Affleck. Can I come see what you’re doing?’ Sometimes people say yes, to my astonishment.”
Tarek retweeted
What’s happening in the AI-powered reverse engineering and decompilation is absolutely insane. Pretty soon all software will be de facto open-source.
AI is coming for everything and everyone.
Tarek retweeted
Talking to Amjad yesterday, I realized Replit is more important than I'd previously thought. It's the way AI pervades organizations. Organizations can become somewhat aligned with AI by using AI-powered apps, but the deepest change takes the form of code.
Tarek retweeted
Full analysis: investmentbrief.ai/p/ai-inve…
Tarek retweeted
🚨 ALL-IN INTERVIEW: Daniel Ek
Thanks to our partners:
@CPIWealth & @northwestagent
- Fixing American Healthcare
- Catching Disease Before It's Too Late?
- Applying the Spotify Playbook to Healthcare
- AI's Upside & Open Source
++ much more!
(0:00) @eldsjal joins the Besties!
(1:14) Starting Spotify: Beating piracy and betting his own money on the record labels
(6:10) Neko Health: The Spotify playbook for healthcare, the $499 body scan, and why US healthcare waits until you're sick
(36:04) AI: Tech's failure to sell the upside, open vs. closed models, and regulating compute
(44:24) Back to Spotify: Podcasting's open standards and the Stardoll origin story
-----------------------------------------
Thanks to our partners for making this possible!
As wealth grows, complexity compounds. Between investments, tax strategy and estate planning, coordinating it all can feel like a full-time job. Creative Planning can help with the heavy lifting. ad.doubleclick.net/ddm/track…
Starting a business? Northwest Registered Agent gives you everything you need to build a complete Business Identity including free tools and built-in privacy. Get more at northwestregisteredagent.com…
Tarek retweeted
When I was a first-time founder, I cold emailed a GP at Sequoia.
At the time, I'd just moved to Silicon Valley with no credentials and no network. That cold email got me a pitch meeting. Same thing happened at Founders Fund and Khosla.
I won't name GPs, mostly because I don't want anyone getting swarmed with cold emails. But this is what I love about Silicon Valley: nobody here is too successful to take a meeting with someone starting out.
Tarek retweeted
Very cool: Robinhood is underwriting the Oura IPO.
Robinhood’s IPO Access business previously just got shares from the banks underwriting an IPO and distributed them to retail.
Now, Robinhood is actually part of the underwriting syndicate alongside (though far smaller than) Goldman, Morgan Stanley, JPMorgan, BofA, etc.
Robinhood is helping sell the deal, taking underwriting risk (meaning they make a firm commitment to buy shares in the IPO), and will earn underwriting fees.
It’s cool because it very much aligns with their goal of democratizing investing… and they have something the bulge bracket banks can’t offer: direct distribution to 10s of millions of retail investors.
Historically you wouldn’t really want retail in the IPO, but RH makes it easier by saying “we’ll take a $100M block” and you don’t have to deal with 200k tiny investors.
Also notable: they were already investors in Oura via RVI, their closed end fund.
They are now investing in seed companies too via RVI 2, they could theoretically be a lifecycle capital partner to companies.
Tarek retweeted
And this is how @PalmerLuckey found the world-class team to build Anduril, just a week after leaving Facebook.
Tarek retweeted
Heard from VC LPs this week:
A valuable/influential LP shared red flags 🚩 their IC watches for in a toppy market 🧐🤔
1) raising the biggest fund possible
2) uncontstrained pacing
3) erratic portfolio construction
4) strategy or sector drift
5) exuberant valuation posture
Tarek retweeted
One of the most important things to know about this AI cycle is that no one knows what they're talking and even very smart and plugged in people are continually reversing their opinions and then re-reversing them three months later.
Tarek retweeted
Two of the biggest VC wins of the year were complete pivots (cursor pivoting from AI CAD and Hugging Face pivoting from a consumer chatbot)
For founders, this is a great lesson to never give up till the money is gone
Tarek retweeted
A YC startup tried adding "…for data centers" to the end of their company description in the YC Directory.
They told me 25 investors reached out the next day.
Wild times
Tarek retweeted
One aspect of the VC GP-LP dynamic that I will never understand is that the industry appears to the structurally set up to reward the least disciplined managers in the short term.
VC: "We deployed our fund in half the time we told you we would. We need more money."
LP: "Well done! Here you go!"
The direct result of this is that the least disciplined GPs are the ones earning the highest fees per unit of time and consuming the largest portion of LP capital.
GPs that deploy on the schedule they set are directly penalized.
Tarek retweeted
The dirty secret of VC rankings is that most of them are designed with an outcome in mind.
Every time a list is released, the reaction from the VC community is to pick it apart for any way that it diverges from expectations.
"X isn't in the top 10? This isn't credible."
"Y is in the top 10? No way this is accurate."
In response, the top places have become a static reflection of "tier 1" firms, regardless of merit. This is achieved by treating scale as a proxy for success.
Occasionally it's possible to reverse engineer scoring and introduce quality of returns (dollar efficiency) into the calculation. Doing so will predictably elevate firms who are known to have exceptional DPI, displacing the agglomerators with weaker returns.
For example, applying a weighted moving average of fund size to the outcomes scored in @IlyaStrebulaev's ranking, produces the following shifts:
1) @usv (+37 places)
2) @firstround (+36 places)
3) @svangel (+30 places)
4) Inflection Ventures (+28 places)
5) @Lux_Capital (+27 places)
6) @atlasventure (+21)
7) @MenloVentures (+20 places)
8) @shv (+18 places)
9) @RibbitCapital (+16 places)
10) @CRV (+15 places)
Resulting in a new top 10, which is:
1) @svangel
2) @RibbitCapital
3) @benchmark
4) @sequoia
5) @kleinerperkins
6) @IndexVentures
7) Parkway Ventures
8) @shv
9) @firstround
10) @global_dst
This update is limited to the list of 50 that the original methodology produced. In reality, a true top 10 by DPI would probably contain many less-well-known microfunds and seed firms.
In reality, the only way to get anything useful out of these rankings is to understood how they are calculated and what that represents. Unfortunately, hardly anyone reads the methodology.
More on this here: blog.joinodin.com/p/ranking-…
