90% of AI 'breakthroughs' are slop. Let's build the real 10%.
Bucharest, Romania
Joined December 2021
- Tweets1.3K
- Following334
- Followers119
- Likes715
Flavius Burca retweeted
Meetup #5 is happening on Nov 4.
We’re continuing the Agentic Engineering series with Loops, graphs & workflows, hosted by @stripe at their Bucharest HQ.
On the menu:
• 2 rounds of live demos
• open mic
If you’ve got something cool to show, this is your sign 👇
Flavius Burca retweeted
Chinese AI company Axera claims AX8860 AI chip can beat the DGX Spark in AI LLMs by 25%, offering 273 GB/s bandwidth, 128 GB LP5X memory, all at 40W, while a 4-module PCIe card pumps out over 1000 TOPs. 🔗 wccf.tech/1ldd1
Everybody's showing off single user decode speed. Who cares ?
Agentic use is 128k+ context. Show me prefill @ 128k context. Show me decode speed @ 2-3 users for a 128k prompt.
Computing prefill attention= O(n2). Decode attn is O(m) and generating tokens with KV cache is O(nm + m2).
Decode speed for single-user simple prompts tell nothing and are misleading !
Always tell the benchmark context: prompt len, engine, version, concurrent users, GPU, model, quant DDR? and OS.
Surogate Rune v3 now matches Jev on all levels !!! The distance is within noise: 57.9 vs 57.4 over a benchmark of 150k decisions covering all domains.
Rune v3 also has SOTA image decision support.
huggingface.co/spaces/multim…
Flavius Burca retweeted
hey @huggingface @pollenrobotics @MicroDuckNVDA check our latest decision model called surogate rune that is 2nd next to jev! look at it controlling microduck in real time!
Flavius Burca retweeted
Surogate Rune is out: an open-weights decision model.
It answers with a decision, not prose: which option, how likely, how it rates on your scale. From text and images.
Top open model on the Decision Index. 42 ms decisions on your own GPU. Apache 2.0.
invergent.ai/blog/rune/
There are 3 levels of AI adoption:
Level 1: Ask ChatGPT questions
Level 2: AI inside existing workflows
Level 3: Re/Build the workflow because AI exists
Most companies are at #1
The beast is training at 25.3k TPS on native fp8 tensor cores. Only with the surogate trainer, nothing else comes even close.
Guess the GPUs :)
github.com/invergent-ai/suro…
Surogate Rune is out: an open-weights decision model.
It answers with a decision, not prose: which option, how likely, how it rates on your scale. From text and images.
Top open model on the Decision Index. 42 ms decisions on your own GPU. Apache 2.0.
invergent.ai/blog/rune/