@madiator is there a way to trace where search requests are going to?
English
Michael Burkov
133 posts
Michael Burkov
@xmikebur
Sales Engineer, AI Inference Platform @nebiustf
New York, NY Katılım Eylül 2023
44 Takip Edilen32 Takipçiler
When CC or Codex do web search, what are they using underneath?
English
I am getting really comfortable with cold calling.
Should I be proud… or deeply concerned? 🤔
English
RL workloads are bursty by nature. Plan capacity to avoid over/under‑provisioning, and make autoscaling a first‑class feature.
English
Batch processing is everywhere. Classic map reduce is a popular approach for working with AI data. Map operation gives you flexibility to run custom code on your data to prepare datasets for training and RL. Cursor team talked a lot about it.
English
Came back from #raysummit by @anyscalecompute . I spoke w/ and learned from AI operators at Notion, Uber, xAI and the like. AI infra trends affecting engineers working in the space:
English
Reliability across nodes matters most. In distributed jobs, hardware fails (not if but when). Building a system that restarts failed jobs in the multi node environment is a complicated engineering task. I heard this from many people.
English
AI native compute is here. In the context of ML, we are used to data processing running on CPUs, model training and serving on GPUs. This now is consolidating around CPU/GPU compute hardware capable of processing various workloads. Ray leads the way here.
English
“The next great VC firms will look like media companies”
Seb Johnson@SebJohnsonUK
Last week @HarryStebbings did a post on LinkedIn for a company which directly led to $3.5m in ARR for them. That is WILD. I don't how other VCs can compete with that. And it doesn't seem that most VCs are even trying to. @a16z are building this level of media distribution with @eriktorenberg at the helm, but most other Tier 1s are just relying on their brand to maintain their lead. I just don't see how that works out long term.
English
OpenAIs cannibalism of the horizontal startups that build on it will be studied in HBS and GSB for decades
Jerry Liu@jerryjliu0
Take: I don’t think n8n is dead OpenAI’s AgentKit is nice but imo it takes a big company’s worth of resources to maintain a 'real' low-code builder with every integration, enterprise controls, customizability, etc. etc. Just look at any RPA company out there. Could OpenAI build this? Yes. Should they though? There’s so many other things to own.
English
@viggy28 makes more sense. I "consume" both of those services - substack and podcasts
English
@xmikebur Good point, but a lot of initial ICP aren't using those.
We see Substack, podcasts, etc.
English
What are the two more knowledge libraries that we should integrate before GA?
English
@robertnishihara @raydistributed @anyscalecompute Felix Heide's twitter links to a different person :)
English
Ray Summit is going to be excellent. Can't wait to hear from xAI, Perplexity, Cursor, Thinking Machines, Physical Intelligence, Applied Intuition, Prime Intellect, vLLM, and so many others.
Some major themes this year:
- Reinforcement learning infra
- Multimodal data (lots of video)
- Distributed inference
- Scalable agent infrastructure
- Robotics / autonomy
English
@codingyash @modal While you wait run serverless GPU workloads on us. We aren't sending them to aws like Modal but process in our own data center. Dynamic scaling, pay as you go.
tracto.ai/pricing - 30% cheaper than Modal
playground cluster with samples at playground.tracto.ai
English
@madhavsinghal_ While you wait run serverless GPU workloads on us. We aren't sending them to aws like Modal but process in our own data center. Dynamic scaling, pay as you go.
tracto.ai/pricing - 30% cheaper than Modal
playground cluster with samples at playground.tracto.ai
English
modal is down and having an incident
first time ever tbh for me, but still
English
@ivanleomk Run serverless GPU workloads here tracto.ai . We aren't sending them to aws but process in our own data center. Dynamic scaling, pay as you go. 30% cheaper than modal.
playground cluster with samples at playground.tracto.ai
English
In the next 3-6 months, we'll start seeing first AI agents running for days and not just hours
English
As foundational models converge in capability, the true edge for AI products comes down to 2 things:
1. How fast you can iterate.
2. How well you can use data to improve model behavior.
Learn from Cursor
cursor.com/blog/tab-rl
English