AI compute isn't just for inference 2/10
Anthropic’s huge compute commitments raise a deeper point: a lot of compute may be more valuable in research, pre-training, and post-training than in inference alone.Inference monetizes today. Research can improve every GPU going forward.
The AI economy, part 1
It looks at why the build-out is no longer just about GPUs: it’s also about power, cooling, networking, construction, and financing. It also introduces a self-reinforcing loop: better AI → more value → more compute → better AI.
#AI#DataCenters
Multi-agent systems need orchestration
More agents ≠ better results.
Expertise can get diluted, communication can add noise, and long-horizon systems can lose context. The harder problem is orchestration: route the right expertise, limit unnecessary communication
Long-horizon agents need layered memory
A long-running agent also needs its own recoverable trajectory: what it tried, what changed, what failed, and what to revisit. ACM + PRO-LONG point toward layered working, episodic, structured-state, and semantic memory.
Long-Horizon Agent Memory Is a Selection Problem
Bigger context windows do not solve long-horizon agent memory. Over hours or
days, the key challenge is deciding what stays detailed, what gets compressed,
what should be retrieved. Memory is selection,
not just summarization.
Unsupervised On-Policy Self-Distillation
No teacher. No environment hint.
Run the model many times on the same problem. If runs disagree, the majority answer becomes the training signal.
many runs → disagreement → consensus → retrain → better policy
Part 3 of our OPSD.
On-Policy Self-Improvement
An agent calls a tool that doesn't exist.
Instead of discarding the failed trajectory, the environment supplies a targeted hint such as “Tool not found.”
trajectory → mistake → hint → training signal → improved policy
Part 2 of our OPSD series.
On-Policy Self-Improvement
Off-policy: a student learns from trajectories generated elsewhere.
On-policy: the student learns from trajectories it generates itself
But if we remove the teacher entirely, where does the improvement signal come from?
Part 1 of a 3-part series
What if the future of AI isn’t one giant model, but a team of smaller specialized models?
Speculative decoding already points in this direction: small models draft, large models verify.
SpecForge, SGLang, vLLM and Bole are making this increasingly practical.
#AI#LLM
Why open source in the AI era
AI makes software cheaper to build and easier to reproduce. That shortens the life of technical differentiation.
One response: open source — trade implementation exclusivity for distribution, feedback, reputation, relationships, and faster learning.
DSpark points toward a practical direction for LLM serving: draft faster and reduce generation latency without replacing the target model.
The same goal shows up in distillation too: smaller + faster, while preserving useful quality.
#DeepSeek#DSpark#LLM#AI
Open-weight models are changing the enterprise AI playbook. Proprietary data can become a fine-tuned internal asset, but training must work across different accelerators and compute platforms. That orchestration layer is what we are building at Brazen Lab.
A chatbot answers. An agent acts. The hard part is making every step reliable: observing systems, choosing tools, taking actions, and evaluating the full trajectory—not just the final answer. #AIagents
I had six AI-generated videos ready to publish and scrapped every one.
The feedback split cleanly: people liked the content, and disliked that my voice was synthetic.
So now it's unrehearsed and unpolished. Same topics: reliable agents, open-source models.
Hiring for an unusual role @mirendil: Technical Chief of Staff, GTM & Ops. I'm looking for someone exceptional, high-energy, and technically fluent who'll work directly with me and touch everything: strategy, GTM & operations, multiplying our impact.
https://t.co/XEyxcCSdHv
If you are passionate about accelerating science, join us!
3/ That's the series: capacity and trust, from an engineer's seat.
Real experiments, real failures, actual code — no product pitch.
New episodes several times a week, sometimes daily.
Everything → https://t.co/TL8r917WD5
2/ Trust.
Give an agent tools and it'll happily follow a malicious instruction it read on a web page.
The work: red-team agents the way a real attacker would, and harden them before someone else finds the hole.
1/ Capacity.
Everyone fights over the same GPUs while older chips, spot fleets, and idle accelerators sit idle.
The work: turn that scattered compute into one usable pool — low overhead, no rewrites, no lock-in.