Developer · Daily · September 30, 2026

firebase sdk meltdown and sdk dependency risk

A bad Firebase iOS SDK pushed in the morning caused widespread crashes with some apps seeing near‑100% session crash rates.[1] Quizlet alone logged over 2,700 crash reports in 25 minutes before people pinned the issue to Firebase.[2]

Devs are openly questioning Google’s release process, citing lack of automated tests/monitoring and comparing this to the old Facebook swizzling incident.[3]

Complaints about Firebase being understaffed, recurring outages, slow incident comms, and heavy dependence for auth/analytics/push are driving active discussion of alternatives.[4]

llm stack and cost: gpt‑6.1 sol, decisions api, codex pro

GPT‑6.1 Sol is marketed as near‑Astra intelligence at about one‑fifth the price, with a 95% cache read discount and roughly 8x speed.[5]

It costs less than a quarter per task vs GPT‑6 Astra, replaced GPT‑6 Sol after 7 days, and scores only 1 point below Astra on the Intelligence Index, while many users say Sol is effectively comparable for coding.[6]

Sol ships with a 1.05M‑token context window and up to 128k‑token outputs, and OpenAI has now scrapped Astra 6.1 entirely over safety issues.[7]

GPT‑6.1 Sol is already wired into GitHub Copilot, boosting multistep coding, while OpenAI’s new Decisions API is reported to beat Jev on decision benchmarks as Jev lags in latency (70–500 ms vs faster models like Liquid d1 and Intern‑Decision).[8]

On the cost side, API users are reporting $20k–$30k monthly bills, prompting OpenAI to reopen the $200 Pro tier with a new usage calc that roughly halves typical Codex API spend and an Ultrafast tier hitting up to 300 tokens/sec or about 8x throughput.[9]

sql, duckdb, and brittle llm‑generated queries

DuckDB, now dubbed the “world’s friendliest SQL database,” is at 40M+ monthly downloads and just got bought by AWS, cementing it in serious analytics stacks.[10]

Tools like dbx act as thin database clients over 100+ backends (MySQL, PostgreSQL, SQLite, etc.), and engines like Quail plus mcp-server-db2i are pushing an AI‑first SQL workflow across many databases.[11]

Real‑world reports say LLM‑generated SQL quietly mis-specs joins and metrics, with guidance to treat “writing SQL” as a tiny slice of data work compared to schema design and validation.[12]

Even in the Microsoft Fabric/Power BI world, practitioners emphasize models are useless without reliable business-contexted data, which clashes with the current trend of letting agents own production queries.[13]

js/npm and mcp security: ai as an attack surface

The npm ecosystem is in a live fire drill: nine npm packages were caught distributing a self‑replicating SSH worm, and fake Express packages are being used to spread Linux worms.[14]

Attackers are slopsquatting fictitious package names hallucinated by AI coding assistants, pre‑registering them so a casual `npm install` on an AI suggestion can fetch malware.[15]

An MCP server now exists to detect these hallucinated names, and MCP security scanners are emerging to plug this gap.[16] At the same time, a serious OAuth bug in the MCP Python SDK could allow account takeover, reinforcing community worries about running arbitrary Python MCP code on production machines.[17]

Stricter npm security changes are also causing bans and friction for legitimate users, pushing more people toward external shields like Cloudflare for security management.[18]

local llm inference: vllm, llama.cpp, and single‑gpu stacks

vLLM added expert RAM offloading to make frontier-scale models more feasible on local or hybrid setups by pushing more state into system memory.[19]

People are using Ray + KubeRay with vLLM to scale inference across GPU nodes, though one report showed 15 of 17 listening sockets opened without being declared in container ports, highlighting operational sharp edges.[20]

On the lean side, llama.cpp is running Qwen 3.8 27B Q4 with a 100k context at about 30 tokens/sec on a single 16 GB RX 7800 XT, and Qwen-family models now support 32 audio model families.[21]

New small models like Mica v0.1 4B can run decision workloads on an 8 GB GPU, with training costs under $30 of GPU time, and there’s an active debate over whether vLLM or llama.cpp gives better real-world performance for these scenarios.[22]