Radar
The releases, frameworks, and ideas that actually changed how we build. Not every paper—just the ones that mattered. Dates are approximate; links go straight to the source.
17 events
Moonshot AI releases Kimi K3 (2.8T) and Thinking Machines releases Inkling (975B MoE), rivaling frontier proprietary models.
Why it matters Closed the perceived gap between US proprietary models and Chinese/indie open-weights.
Anthropic releases Fable 5, dominating coding benchmarks and correctness.
Why it matters Cemented Anthropic as the leader for reliable, high-stakes knowledge work.
OpenAI releases Sol, Terra, and Luna, bringing extreme inference scaling and dynamic reasoning modes.
Why it matters Formalized the 'test-time compute' trend as a tiered product offering.
The industry definitively shifts from scaling pre-training to scaling test-time compute for reasoning tasks.
Why it matters Cost and latency tradeoffs became the primary architectural challenge.
High-capability reasoning model that rattled pricing and open-research assumptions.
Why it matters Showed frontier-ish reasoning can arrive from unexpected cost curves.
Models trained to spend more compute on internal chain-of-thought style reasoning.
Why it matters Shifted product focus from “chat fluency” to deliberate problem solving.
Omni model with tighter realtime multimodal interaction.
Why it matters Normalized voice/vision as default LLM interfaces, not add-ons.
Stronger open-weight models that keep the open ecosystem competitive.
Why it matters Sustained pressure on closed APIs for cost-sensitive deployments.
Anthropic’s Opus / Sonnet / Haiku set a new bar for helpful, long-context assistants.
Why it matters Intensified competition on reasoning quality and developer UX.
Sparse Mixture-of-Experts models deliver strong quality at lower active compute.
Why it matters Made efficient scaling a first-class architecture choice again.
Function calling, browser tools, and multi-step planners become standard LLM product patterns.
Why it matters Moved LLMs from “answer boxes” to systems that act in software.
Multimodal frontier model with stronger reasoning and broader usefulness.
Why it matters Set the commercial bar for general-purpose AI assistants.
Meta releases capable open-weight LLMs that ignite the open model ecosystem.
Why it matters Made serious LLM work possible outside closed API vendors.
RLHF’d conversational interface makes foundation models mainstream.
Why it matters The product moment that pulled LLMs into every industry conversation.
Open latent diffusion model brings high-quality image generation to consumer GPUs.
Why it matters Democratized generative image models outside closed APIs.
Self-attention replaces recurrence as the dominant sequence architecture.
Why it matters The architectural root of nearly every modern LLM.
GPU-trained CNN wins ImageNet by a huge margin - deep learning’s public breakout.
Why it matters Proved scale + GPUs beat hand-crafted features for vision.