AI Blog
AI Blog
The latest AI trends, enterprise adoption tips, service use, seminars, and talent development — insights to support decisions at both the management and frontline levels.
Categories
Latest posts
AMD対NVIDIA ― 自社でAIを持つなら、どちらのGPUか
「GPUはNVIDIA一択」はもう古い。学習はCUDA優位、推論はROCmが実用域に。B200とMI350系、VRAMと価格で選ぶ小規模構成、自社保有とクラウドの損益分岐まで、オンプレAIの実務目線で整理します。
Read moreIT属人化の処方箋 ― AI同士が引き継ぐ「SmartOptimizer」
「あの人がいないと分からない」を仕組みで解く。AIを入れてもAIごとに記憶が閉じれば属人化は消えません。AI同士が要約を引き継ぐ共有メモリで、退職・異動でノウハウが消えない体制をつくります。
Read moreWill AI Cost Less in the Future, or More?
The price per token keeps falling, yet AI bills keep growing. We look at the structure that raises the amount of work done per request — reasoning tokens and high-resolution image input — and why today's monthly plans may not stay on today's terms.
Read more[Project Diary ⑦] Dropping the paper comparison and deciding by measuring 10,000 records
Performance settled in requirements definition cannot be held by a paper comparison table. We stood up a verification environment, loaded 10,000 records, and confirmed first that no screen worked. From 22 seconds to under 3, and on to finding a hole in the design through an incident.
Read more[Project Diary ⑥] Cutting promises we could not keep out of the material
Promises written with good intentions creep into proposal material. A reporting frequency heavier than the contract, operating rules that will not hold in reality, words that put the other side on guard. The record of cutting them out before the kickoff.
Read more[Project Diary ⑤] Killing “it looks like it worked but nothing happened”
The worst-behaved thing in a system is processing that fails silently without raising an error. A success message appears but nothing is saved; a list is quietly cut off partway. The record of combing out and killing those quiet failures.
Read more[Project Diary ④] Bringing a demo prototype closer to the real build
Passing a demo and surviving being rebuilt are two different things. The record of matching how data is held to the customer’s current setup, replacing apparent processing with the real thing, and pushing conditions written into the screen out into settings.
Read more[Project Diary ③] Counting what we did not understand, up to 37 items, and putting it on the table
Requirement interviews spin in place because the questions are abstract. With something that works in front of you, it sharpens to “this field on this screen cannot be filled in.” The record of counting 37 unknowns and extracting requirements from two directions.
Read more[Project Diary ②] Cutting our own company introduction out of the first meeting’s agenda
When something that works exists before the first meeting, the moves available in sales change. Compressing the self-introduction to secure 15 minutes of listening, lining up 16 awkward confirmations first, deliberately holding the demo to five minutes. The record of what AI-driven development brings to the early stage.
Read more[Project Diary ①] 82 fixes to the prototype on the day we won the project
AI-driven development is not only about the implementation phase. Everything from sales through to kickoff changes. What we did on the first day was not reading documents but fixing 82 things in a working prototype.
Read more[Field Report ⑥] Rebuilding a stalled Android app with AI, signing key and all
An Android app whose push notifications had not arrived for years, rebuilt natively with AI while keeping the signing key issued in 2016. The dead delivery platform was moved to FCM, restoring notifications to roughly 3,200 devices.
Read more[Field Report ⑦] Reviving a stalled iOS app by rebuilding it in Swift with AI
An iOS app silenced by a previous-generation notification route, rebuilt with AI on Swift + WKWebView. Notifications were restored to roughly 10,800 devices — the majority of members — without replacing the old server.
Read more[Field Report ⑤] Rescuing a 10-year-neglected in-house system off an end-of-support server in two days
A 2016 system left running on an OS that reached end-of-support in 2020. This is the record of migrating 3,623 tickets and 4.0 GB of attachments to a modern environment with zero data loss. Of the eight problems we hit, only three came from the application itself.
Read more[Field Report ④] How an old mining PC gathering dust in storage became an in-house "private ChatGPT" in two days
Four used GPUs, almost zero extra spend. This is the record of running a GPT-4o-class 30B model on outdated hardware without CUDA and building a company-wide private AI platform in two working days — automatic failure recovery included.
Read more22 systems and 246k lines on a single PC — a measured record of 3.5 months of AI-agent development
One engineer + an AI agent + one PC built 22 systems and about 246,000 lines in 3.5 months. We share the measured figures tallied from Git history and a breakdown of the equivalent 10–20 person-years of effort.
Read more[Field Report ①] Fully modernizing a 20-year-old Java system in about a week with an AI agent
Java 1.7, no documentation, no original developer. The record of using AI to analyze an "untouchable legacy," then rebuilding every member-facing feature plus 18 admin screens in about a week — with the DB schema left in place.
Read more[Field Report ②] Adding a new entry point to a core system without changing a single line of existing code
A strangler-pattern record: adding a thin REST layer and a new gateway (SSO auth) to a core system running on an end-of-support OS, without modifying existing classes — and passing production E2E in just days.
Read more[Field Report ③] Wrapping up a month-long architecture study with "a working prototype in 2 days + real measurements"
What used to take about a month of desk-bound comparison tables: with an AI agent we built a ~66,000-line prototype in two working days and, after measuring 10,000 records to confirm every API dropped from 22s to under 3s, locked in the architecture.
Read moreWe'll be sharing the latest AI insights here.
We share the latest AI trends, enterprise adoption tips, and insights on security and talent development — clearly, from both the management and frontline perspectives.
Read more