Want to build something crazy in 1 prompt?
Let's meet at Shift:
- ⏳ 48 hours of hacking
- ✨ unlimited Codex+Cursor credits, awesome coaches
- ⏰ 27-29 March, Nantes
shift-hackathon.com
Token Meter for VSCode — free, open source.
Works on Cursor too.
🔗 github.com/samber/vscode-token-meter
⭐ if it's useful.
Writing code isn't the bottleneck anymore.
Distribution is. Compliance is. Paperwork is.
We spent 10 years automating the build. Nobody automated the publish.
The "long" part:
- Creating an Azure Publisher account
- Filling in marketplace metadata
- Waiting for review
2 hours of bureaucracy for a tool that took 1 prompt to exist. 🤮
So I opened Claude Code and described what I wanted.
A VSCode extension. Token count in the status bar. Real-time. Switches model on click.
One prompt. It built it.
I was building Claude Code Skills: custom instruction sets injected into the AI context window.
Problem: no idea how many tokens each skill was consuming.
Was I burning half the context budget on docs? No clue.
The longest part of my last side project was filling out a form.
1 prompt to build. 2 hours to publish.
We're optimizing the wrong thing.
🧵
Your Go code is leaving 90% of the CPU idle ...until now.
samuelberthe.substack.com/p/your-go-co...
samber/ro, the streaming alternative to samber/lo, also supports SIMD operations in its last version !
Changelog: github.com/samber/ro/re...
Doc: ro.samber.dev/docs/plugins...
SIMD documentation is available here: lo.samber.dev/docs/experim...
(requires a recent amd64 CPU)
🎉 We just made the biggest release of samber/lo to date: github.com/samber/lo/re...
- 🧪 adding support for SIMD (Go >= 1.26)
- 💥 adding **Err variants helpers
- 🌊 lots of performance improvements
The library has just surpassed 1.000 helpers!
These GPU/LPUs are still niche, but they’ll be game-changing for near-real-time use cases.
* Their direct competitor, Groq, has very little SRAM per chip. Running a 70B model requires assembling roughly 1k LPUs in a large cluster, which introduces additional latency.
* Memory capacity per chip is unknown, but it matters because batching relies on abundant memory.
Cerebras can achieve almost the same token bandwidth with batching (around 2k tok/s per user, up to 20k tok/s per chip).
* In the future, chips without HBM will probably be paired with traditional Nvidia GPUs: B200 for prefill, followed by Groq/Cerebras/Taalas for decoding.
17k tok/s is impressive. A few observations:
* This measures only the decoding stage. We have no information about the prefill (Time To First Token)
* Taalas HC1 likely has lower TFLOPS than the B200, so the performance gap during the prefill stage might not be significant.
👇🧵
"I Verified My LinkedIn Identity. Here's What I Actually Handed Over."
thelocalstack.eu/posts/linked...
🧠 Can someone explain what happened in just one weekend? Clawdbot has been around for 5 months but has only just exploded in popularity.
Probably the fastest product launch ever on GitHub. 🎸
👀
100 people coding in a single repository is a merge nightmare.
If you have 100 *AI-assisted* coders in a single repository, you'd better have a f****ng fast CI.
A picture is worth a thousand words.
The production of the Blackwell GPU generation starting 2025-H2 has caused DRAM prices to soar.
If this technology is real, this is a ChatGPT-moment.
*They have 0 factory, industry takes time.
*They did not disclose the materials
#CES #battery 👇
www.youtube.com/watch?v=Y-aP...
China successfully reverse-engineered an old ASML machine. 😳
(it's a prototype - no working chip has been produced ...yet)
www.reuters.com/world/china/...
👾 🎧 If you like coding and music, look at that: www.youtube.com/watch?v=iu5r...
Seems like Cloudflare still uses their old "FL1" proxy (based on Openresty and Lua) for at least 25% of total traffic. 🤯
Even the leaders struggle to eliminate their technical debt.
blog.cloudflare.com/5-december-2...
How Meta keeps its AI hardware reliable 👇
// Training a model on 10k-ish GPUs will toast a few hundred along the way. Better make sure you catch silent failures 😅
engineering.fb.com/2025/07/22/d...
🧠 Claude is becoming the default model for developers, for both AI-assisted coders and agent creators.
Look at that 👇 #contextEngineering
www.anthropic.com/engineering/...
Deepseek, October 2025:
"Hey guys, look at our new deepseek-ocr model!"
Deepseek CTO in 6 months 👇
1) ourworldindata.org/data-insight...
2) ourworldindata.org/reducing-fer...
