flashnik @flashnik

Hacker and Entrepreneur in AI. Building voice-first AI agents for real businesses. linkedin.com/in/flashnik Joined February 2008

Tweets

9
Followers

47
Following

41
Likes

1

flashnik @flashnik

2 weeks ago

My bet: Qwen-with-a-head, ModernBERT and today’s diffusion approaches are points in a search space, not the final architecture. I wrote up the architectures, benchmarks, calibration, serving and production trade-offs: agentunicorn.ai/research/jev-a…

0 0 0 1 0

View Details

flashnik @flashnik

2 weeks ago

I spent the last week going through #Jev by TypeSafe and the open implementations around it. I expected a story about cheaper classification. Instead I found an architecture search across causal LLMs, encoders, NLI, calibration-first models, diffusion and more

7 0 1 11 0

View Details

flashnik @flashnik

2 weeks ago

The business opportunity may be larger than replacing existing LLM calls with cheaper ones. If the marginal cost of another semantic decision falls enough, we can automate long-tail decisions that today stay rule-based, manual — or simply never get implemented.

0 0 0 1 0

View Details

flashnik @flashnik

2 weeks ago

This is where the architectural shift happens. Instead of asking one generative model to do everything, use the right inference primitive for each step: deterministic rules, retrieval, specialized classifiers, decision models — and generation where generation is actually needed.

0 0 0 1 0

View Details

flashnik @flashnik

2 weeks ago

Decision models also do not eliminate ML engineering. You still need representative eval data, traces, calibration on your traffic, drift monitoring, versioned thresholds, formulation tests, fallback policies and replay.A general model ≠ no ML lifecycle.

0 0 0 1 0

View Details

flashnik @flashnik

2 weeks ago

The harder problem is not getting a fast argmax. It is getting a decision primitive that transfers across domains, languages, state formulations and option sets — while returning probabilities useful enough to drive real business actions.

0 0 0 1 0

View Details

flashnik @flashnik

2 weeks ago

The interesting part is not only “don’t generate tokens.” Jev says it processes shared state once and evaluates multiple questions in parallel. Open projects are exploring different ways to reproduce that: shared KV, isolated branches, direct logits, full-option scoring and more.

0 0 0 2 0

View Details

flashnik @flashnik

2 weeks ago

A lot of agent logic today is basically: state → large LLM → JSON → parser even when the actual result is one enum, boolean or probability. Tool routing, RAG/no-RAG, escalation, policy checks, model selection, workflow transitions...