#llm

Bar chart of TypeScript errors at each stage of upgrading an AI SDK 6 app to version 7, with the codemod stage still failing at runtime

AI SDK 7 upgrade: the codemod compiles, but a tool breaks

We upgraded a 55-line AI SDK 6 chat app to version 7.0.148 and ran the official codemod. TypeScript went from 4 errors to 1, then to 0 after one manual fix, but two things still changed at runtime

Thu Oct 22 2026 · 10 min read · 1 views

Go encoding/json v1 and v2 results for eleven LLM-style JSON payloads on Go 1.26.9 and 1.27.2

Does Go 1.27's json v2 break your LLM parsing? We tested

Go 1.27 runs encoding/json on the v2 engine. We tested 11 LLM-style payloads on Go 1.26.9 and 1.27.2: v1 behaved the same, v2 is stricter, Unmarshal is 2.3x faster.

Sat Oct 17 2026 · 10 min read · 2 views

Software

Bar chart style comparison of token counts across languages

LLM token cost by language: we measured 14 of them

LLM token cost by language, measured: the same sentences cost 7% to 54% more tokens in German, Polish or Hindi than in English on GPT-4o's tokenizer.

Wed Oct 14 2026 · 9 min read · 86 views

Illustration of a circular loop with a dot at its center

AI Agents for Laravel Developers: Build One in 42 Lines

AI agents for Laravel developers, explained by building one in 42 lines of PHP. The loop, the parts, a measured cost curve and when not to build an agent.

Fri Oct 09 2026 · 13 min read · 2 views

AI Software

Code editor showing an AI SDK 7 result object with usage tokens aggregated across multiple agent steps

AI SDK 7 Changed usage. Your Token Counts Are Now Wrong

If a project upgraded to AI SDK 7 and its per-request token costs suddenly look higher than before, nothing is actually costing more. result.usage used to report only the final step's tokens. In AI

Tue Sep 29 2026 · 6 min read · 6 views

Claude Opus 5 API request parameters that now return a 400 error

Claude Opus 5 returns 400 on temperature. What to do

Set temperature=0.7 on Claude Opus 5 and the request fails. The API returns a 400 error, not a warning and not a silently ignored parameter. The same happens with topp, topk, and with the manual

Fri Aug 21 2026 · 6 min read · 77 views

An AI agent's context window filling with stale tool output across a long task

Your AI agent gets worse long before the context fills

Agents degrade as context grows, well before the window is full. Here is the token budget that fixes it, and what each fix costs you in latency.

Thu Aug 20 2026 · 5 min read · 10 views

Upgrading a Python agent codebase from Pydantic AI v1 to v2

Pydantic AI v2 migration: the change that throws no error

The riskiest Pydantic AI v2 migration change raises no exception: openai: model names now hit the Responses API. Here is the safe upgrade path.

Thu Aug 20 2026 · 5 min read · 9 views

Software

An AI agent repeating the same tool call because tool_choice stayed forced

Your agent loops forever. It is probably tool_choice.

Your agent calls the same tool repeatedly and never returns an answer. The advice you will find first is to set maxiterations, which caps your bill without addressing the underlying defect.

Wed Aug 19 2026 · 6 min read · 12 views

AI Software

Timeline from the August 3 Qwen 3.8 Max launch to the August 9 weight upload and the August 24 Flash-Next release

Qwen 3.8 Max promised open weights. Here is what shipped

Qwen 3.8 weights are on Hugging Face: the 27B under Apache 2.0 and the 2.4T under a custom license. Sizes, license terms and what you can actually run.

Thu Aug 06 2026 · 7 min read · 12 views