←
Heinrich Wendel1,062 posts
Heinrich Wendel avatar

Heinrich Wendel

@hmw147

south-east asia, tech, health

◷ Joined June 2011▧ 18 Photos and videos
Marсo Reed avatar
⟳ Heinrich Wendel retweeted
Marсo Reed@marco__reed·Nov 3

$MSFT backlog: $677B $ORCL backlog: $664B $MSFT market cap: $3.8T $ORCL market cap: $370B Almost the same contracted backlog. 10x valuation difference. Either $MSFT is extremely overvalued… or $ORCL is the most obvious AI mispricing of the cycle. Which one is it?

Post attachment
Heinrich Wendel avatar
Heinrich Wendel@hmw147·Nov 2

AI has become so good at writing software that the bottleneck is now defining clear requirements and keeping them up to date. Storing them as Markdown files alongside code works, but feels very developer-centric. How do you approach this?

Heinrich Wendel avatar
Heinrich Wendel@hmw147·Jul 22

My thoughts on Gemini 3.6: Either Google screwed up with their pricing strategy, or they have so much demand that they can't catch up with infrastructure.

Heinrich Wendel avatar
Heinrich Wendel@hmw147·May 20

This is big for self hosting of OSS models

KV cache shouldn't disappear every time vLLM restarts. With @novita_labs, we're sharing PegaFlow — a production-grade external KV cache service that plugs into vLLM through the external KV connector interface. PegaFlow runs as a standalone Rust daemon owning the host KV pool, SSD cache, and RDMA resources. vLLM workers attach via CUDA IPC + gRPC, and cache survives engine crashes, upgrades, and model switches. In production-oriented evaluations: 🚀 2.15× faster vLLM startup with a pre-warmed 500 GiB host pool 📈 56% higher throughput for 8 Qwen3-8B instances sharing one cache ⚡ 72% higher throughput for DeepSeek-V3.2 MLA TP8 (logical KV stored once, not per rank) 🌐 194 GB/s average remote-read throughput across nodes Three-level hierarchy: pinned DRAM, remote DRAM over RDMA, local SSD on io_uring. Integrates through the existing `kv_transfer_config` path — no vLLM source changes. 📖 vllm.ai/blog/2026-05-18-pega…

Heinrich Wendel avatar
Heinrich Wendel@hmw147·May 1

I tried @zeddotdev today. Couldn't get used to it (yet): 1/ Default fonts feel way too big with too much tracking, not optimiezd for code 2/ Claude code in agent panel pauses for long times and doesn't seem to finish; can't connect my Codex Pro subscription 3/ Search bar layout breaks all the time, fits about five characters before overflowing into the next line Love the direction of more simplicity, but might need another month or so of polishing.

Heinrich Wendel avatar
Heinrich Wendel@hmw147·Apr 29

Seems OpenResponses API has largely failed to become the universal layer. Everyone still defaults to /chat/completions (with some minor variations). Am I missing something?

Heinrich Wendel avatar
Heinrich Wendel@hmw147·Apr 28

This in itself would boost the S&P 500 by 5-6%.

The biggest IPO run in the history of the market three $1T+ companies. possibly all going public in the next 12 months. SpaceX IPO: $1.75T. OpenAI IPO: $1T Anthropic IPO: $1T we're living through the greatest technological wealth creation in history.

Post attachment
Heinrich Wendel avatar
Heinrich Wendel@hmw147·Apr 27

Smart. AWS distribution of OpenAI GPT models will outgrow any lock-in to Azure. Win-win.

OPENAI AND MICROSOFT $MSFT JUST ANNOUNCED AN AMENDED PARTNERSHIP AGREEMENT Here's what changes: Microsoft will no longer pay a revenue share to OpenAI OpenAI can now serve its products to customers across any cloud provider Microsoft continues to participate in OpenAI's growth as a major shareholder and retains a license to OpenAI IP through 2032 Revenue share payments to Microsoft continue through 2030 (Source Reuters)

Post attachment
Heinrich Wendel avatar
Heinrich Wendel@hmw147·Apr 27

How many data centers could you build for that, powering the next generation of Malaysia - when the fuel runs out?

Malaysia is spending about RM8.28 million per hour, or RM2,300 per second, on fuel subsidies to cushion the impact of global supply disruptions. The surge in crude oil prices has significantly increased fiscal pressure on the government, reports Berita Harian. nst.com.my/news/nation/2026/… 🧵1

Post attachment
Heinrich Wendel avatar
Heinrich Wendel@hmw147·Apr 26

I love taking long range buses in Malaysia for exactly the same reason.

Yesterday I drove my @tesla 900 miles on FSD from Miami to Nashville and I realized it’s genuinely the better option. I fly that route 2 to 3 times a month. Flights are never under $400. Most times $600. Sometimes $800. Add Uber to and from both airports, or parking garage fees. Then factor in the delays, the cancellations, the security theater, the chaos, the guy next to you who hasn’t met deodorant yet. On the other hand: I pack healthy snacks, press one button, and the car just goes. I took calls. Replied to emails. FaceTimed my family. Ate without pulling over. Did everything I normally do on a travel day, except none of the stuff that makes travel days miserable. My biggest concern going in was range and charging. Here’s what actually happened: My bladder needed one extra stop the car didn’t even suggest. Most charging stops were under five minutes. Total cost for the whole trip was less than just the uber to the airport. And this was the base model Y. Now I’m thinking I should get something comfier and just make this the default.

Heinrich Wendel avatar
Heinrich Wendel@hmw147·Apr 26

Love the idea of @__tinygrad__ . Can you put the specs in some more relatable terms. Like: 1/ What’s the biggest model that it can run? (DeepSeek V4?) 2/ How many tokens/s can it produce with a common OSS model (Qwen 3.6 27B?). 3/ How many users can it serve in parallel at that speed? 4/ If I run this at 60/70/80% utilization, what’s my effective token price? 5/ What will be the real intelligence score when running at FP4/8 quantization?

Heinrich Wendel avatar
Heinrich Wendel@hmw147·Apr 21

A few provider fixes that matter when using Agents at scale: github.com/vercel/ai/issues/… github.com/vercel/ai/issues/… github.com/vercel/ai/issues/…