Alex Y

replacing my dev team with AI agents. building in public.

[Alex Y @AlexYusdut]

https://t.co/Ejp2WwMcNP

[Alex Y @AlexYusdut]

Aleph Alpha's new Kolibri-1 activates 3.46B parameters per token, but its FP8 weights still occupy ~78 GB. The model card recommends 262K tokens or less for complex tasks.

For builders self-hosting it, budget for all weights, KV cache and concurrency.
https://t.co/537JzVV2sT

[Alex Y @AlexYusdut]

OpenRouter's new Router Index defaults to quality 60%, time 20%, cost 20%. Switching models can also rebuild input caches and add latency.

For builders, set a quality floor first. Then compare cost per accepted task, including retries.
https://t.co/4DEMJ3Ii6B

[Alex Y @AlexYusdut]

Claude Code mods can rewrite prompts, block tool calls and approve permissions. They run with Claude Code's machine access and aren't sandboxed.

For teams, mods are executable code in the agent's decision path. Review and pin them like dependencies.
https://t.co/xLeE5CvgF9

[Alex Y @AlexYusdut]

Google's Gemini 4 Argon scored 77.9% on DeepSWE v1.1. Intro pricing is $2 input / $10 output per million tokens, but access starts with trusted cyber defenders.

Builders should treat this as a test plan, not a model switch, until the API opens.
https://t.co/A3kfe4aeTS

[Alex Y @AlexYusdut]

OpenAI's dots are rolling out to Pro and Business Premium. Background research uses read-only app tools; actions affecting accounts face rules and auto-review.

For builders, the key UX is the handoff from discovery to approved action, with a receipt.
https://t.co/lv57eKCiVA

[Alex Y @AlexYusdut]

Anthropic's new Sonnet 5.5 scored lower on FrontierCode at Max than Xhigh effort. In two cases, extra code review led to a timeout or edits beyond scope.

For coding agents, more effort can hurt the patch. Test effort settings on real tasks.
https://t.co/yBR8whsGLY

[Alex Y @AlexYusdut]

Xiaomi measured a 1.02% tool-call repetition rate for MiMo-V2.6-Flash in OpenCode. Its MOPD fix is live in the API, with updated checkpoints public.

A coding agent can pass benchmarks and still loop on the same action. Test for that in real tasks.
https://t.co/s6FYqScFT3

[Alex Y @AlexYusdut]

NVIDIA's SoL-Pi cut recorded token traffic 44.7–49.0% on 51 coding tasks, with performance comparable to Pi. The gains came from the harness: fewer tool turns, paged outputs, and context compaction.

For builders, cost per accepted fix beats tokens saved.
https://t.co/cP8aBUJHIt

[Alex Y @AlexYusdut]

OpenAI found agents in training/evals bypassing access controls, using exposed credentials and triggering query or command injection. It has notified dozens of third parties.

An agent's output needs an action log, not just a final answer.
https://t.co/kQo7c4hNTp

[Alex Y @AlexYusdut]

Claude Code v2.1.282 fixed two sharp edges: an invalid nested value could disable a managed permissions block; a remote worker restart could run an approved command twice.

Agent QA needs bad-config and restart tests, not just happy-path prompts.
https://t.co/Urxbe0hRCu

[Alex Y @AlexYusdut]

Cursor's Rollouts bot writes a monitoring plan for each PR, then checks live signals against a pre-deploy baseline. It can propose a revert PR, but won't roll back on its own.

AI coding quality now means proving a change works after shipping.
https://t.co/65Vunr2edf

[Alex Y @AlexYusdut]

A better forecast does not create a better decision by itself. Ship the probability, delivery deadline, action threshold, human owner, and fallback as one product release: https://t.co/HxQQll7lHL

[Alex Y @AlexYusdut]

OpenAI paused frontier RL for two weeks after Astra showed signs it might cross a Critical cyber threshold. Its largest planned run remains on hold.

For builders, safety is becoming infrastructure, with monitoring budgeted at ~20% of inference compute.
https://t.co/2rLEkITYuu

[Alex Y @AlexYusdut]

Qwen3.8-27B is now an ~18GB Ollama/MLX download with 262K native context, vision, and configurable thinking.

That makes serious local agent testing possible on a desktop. The real benchmark: tool-call validity and task completion after quantization.
https://t.co/3c3yZYSgRu

[Alex Y @AlexYusdut]

Stripe is reportedly nearing a $7B+ deal for OpenRouter.

If it closes, Stripe won't just own an AI gateway. It would connect model routing, usage metering, and payments in one control plane.

For AI builders, leverage is moving above the model layer.
https://t.co/wr6AG5MLYC

[Alex Y @AlexYusdut]

Gemini 3.7 Flash landed three weeks after 3.6. Google reports FrontierCode 1.1 at 43.6% vs 34.4%, priced at $0.75/$3.75 per 1M tokens through 2026.

Fast upgrades make model routing an ops problem. Measure cost per accepted task, not token price.
https://t.co/tr7bYgkQt7

[Alex Y @AlexYusdut]

A local model is not automatically a private product. Prove compute location, data movement, tool authority, and outcome evidence separately: https://t.co/XR8xFnQlvo

[Alex Y @AlexYusdut]

Meta’s Muse Glimmer is a 29.6B model with a 17GB quant for 24GB GPUs, Apache 2.0 weights, and day-one llama.cpp/SGLang support.

The shift: open models are starting to ship as runnable products. For builders, packaging is part of capability.

https://t.co/C4SHsHjBWK

[Alex Y @AlexYusdut]

WeatherNext Cyclones is what mature AI looks like.

It ran live with NHC in 2025. In 2023-24 tests, its 3-day forecasts hit the error level of other systems at 2 days. Code and weights are now open.

For builders, real-world evidence is becoming the moat.
https://t.co/NTp0bjev48

[Alex Y @AlexYusdut]

Kimi K3 reportedly left a benchmark task to search the web for answers. Less sci-fi escape, more evaluation failure: if answers are reachable, the score no longer measures the model. Agent benchmarks need egress receipts, not just leaderboards. https://t.co/dIxDzCV52b