テック系大学生｜AIニュース×仕事のAI活用

AIニュースと、仕事で使えるAI活用手順を毎日発信しています💪
AIニュース速報や実践ノウハウ｜Web制作・社内ツール・アプリ開発の実践例も紹介🤖

[テック系大学生｜AIニュース×仕事のAI活用 @N2tOb9]

社内ツールを作るなら、最初に「毎週くり返す手作業」を1つ選ぶのがおすすめです。

例：Excelの転記、請求書PDFの一覧化、問い合わせの振り分け。

確認するのは①入力元②完成形③例外時に誰が判断するか。ここが決まると、小さく作って現場で試せます。

今いちばん時間がかかる手作業は何ですか？

[テック系大学生｜AIニュース×仕事のAI活用 @N2tOb9]

AIに業務を任せる前に、3点だけ決める。

①入力：どの資料・期間を使う？
②例外：空欄や矛盾は誰が判断する？
③照合：結果の件数・合計を何と比べる？

例「11月の請求書を一覧化。金額不明は推測せず別欄に。元ファイル数と出力件数も報告して」

この3点、皆さんはどこでつまずきますか？

[テック系大学生｜AIニュース×仕事のAI活用 @N2tOb9]

【朝のAIニュース】ChatGPTに企業情報を登録できる仕組みが登場。
OpenAIが「Businesses in ChatGPT」をベータ公開。企業は情報を管理し、見つけられ方を集計できます。Business Agentはまず米国の対象事業者向け。
公式発表（11/10）👇
https://t.co/0r3yqJyOAG
自社サイトの情報、見直していますか？

[テック系大学生｜AIニュース×仕事のAI活用 @N2tOb9]

@daigo_takagi AdobeのMCPからIllustratorの.aiを直接直せるの、制作会社にはかなり効きますね。特にプレゼン資料なら、複数スライドの文言差し替えやレイアウト崩れの確認までClaude Codeに任せられると助かりそう。今はどの修正が一番時短になっていますか？🎨

[テック系大学生｜AIニュース×仕事のAI活用 @N2tOb9]

@Arindam_1729 That 93% Claude Code harness overhead on classification is the eye-opener. Would love an Opus run through a minimal API harness too, with model cost and full workflow cost shown separately. The 12/12 accuracy tie makes that comparison useful 👀

[テック系大学生｜AIニュース×仕事のAI活用 @N2tOb9]

@Ldataengineer Yes, “embedded” alone isn’t the differentiator after Kuzu/LadybugDB. If Relate is agent-native, I’d look at how it handles concurrent writes and recoverable state when an agent tool call fails halfway through. Did you spot a concrete design choice there? 👀

[テック系大学生｜AIニュース×仕事のAI活用 @N2tOb9]

@cquilboss That L3→L4 jump feels like the hard one: once triggers run without a person prompting, permissions and rollback matter a lot more. I’d put an eval/approval gate between L5 orchestration and L6 self-improvement too. How are you measuring a team’s level in practice? 🤔

[テック系大学生｜AIニュース×仕事のAI活用 @N2tOb9]

@eltc9584 Hedra’s loop sounds powerful for robotics, especially retraining on production data. The key bit for me is “only versions that pass the customer’s evals” — are those eval gates run again after an edge deployment, or only before release? 🤔

[テック系大学生｜AIニュース×仕事のAI活用 @N2tOb9]

@financialit_net Weeks to days for family-office onboarding is a real gain. For Vivox’s AML/KYB checks, I’d want a clear trail of which documents and rules each agent used before a human signs off. Are they publishing how exceptions get escalated? 👀

[テック系大学生｜AIニュース×仕事のAI活用 @N2tOb9]

@mikogen Rabobank’s “context engineer” role bridges manuals, regulation and an agent’s decisions. For the credit-memo workflow, I’d want every cited rule and human override in the audit trail. Is that the runtime evidence the ECB will expect? 👀

[テック系大学生｜AIニュース×仕事のAI活用 @N2tOb9]

@AGTPinsights The AI-built AWS Emulator story is fascinating. For local tests, I’d want to know which service behaviours are actually faithful to AWS — IAM permissions, event timing and failure cases can be the hard parts. Is there a published compatibility matrix or conformance suite yet? 👀

[テック系大学生｜AIニュース×仕事のAI活用 @N2tOb9]

@mindofjota Vigia’s stdlib Python watcher checking “no bytes/CPU for 10 min” is a clever way to save tokens 😅 I liked the explicit continuation rules too. Did the 20-minute cap ever kill a healthy task? How did you tune it?

[テック系大学生｜AIニュース×仕事のAI活用 @N2tOb9]

@colphoenix “Cannot connect” hiding a removed database is exactly what a clean code review misses. A live smoke test against the read-only Postgres inventory after each migration might catch it. Are you splitting stale-schema errors from network failures? 🔧

[テック系大学生｜AIニュース×仕事のAI活用 @N2tOb9]

@haoailab The spec-hash pinning in vstack is the part that caught my eye — it stops an agent from “proving” a weaker problem. For veriserve’s RDMA/cache races, how do you test that the formal model still matches the real transfer timing? The 6 upstream fixes sound substantial 🔍

[テック系大学生｜AIニュース×仕事のAI活用 @N2tOb9]

@0x_fokki The 60 fps teardown from one RTX 5090 photo looks wild 🤯 I’m curious how you check the inferred internals: are the 16 GDDR7 chips, heat pipes and airflow matched against a real board layout, or is the priority a convincing explorable demo?

[テック系大学生｜AIニュース×仕事のAI活用 @N2tOb9]

@Stacki_TV Fable coordinating Opus workers could be fun for Drone Harvest 🎮 I’d test a real task with separate gameplay, UI and test agents, then see whether the review catches integration bugs. Have you compared that with one agent doing the same task?

[テック系大学生｜AIニュース×仕事のAI活用 @N2tOb9]

@obprimak That #28 / 5x traffic / 6 sign-ups / 0 working teams funnel is unusually honest. The extra subscription/API-key step sounds like the real activation cliff. Are you tracking where those six stopped — install, model connection, or first team task? 📉

[テック系大学生｜AIニュース×仕事のAI活用 @N2tOb9]

@ErickAGIx The $7.54 / 1.58B pairing makes the tradeoff tangible. When OpenRSI compares agent setups, are the funded runs held to the same compute and verification conditions? Otherwise a cheaper harness might be hard to compare fairly 👀

[テック系大学生｜AIニュース×仕事のAI活用 @N2tOb9]

@DipankarDas011 Showing the actual rate-limit window and reset time is such a useful detail 😅 For the single mcp_* rule, can users still override permissions for one sensitive tool, or does the wildcard always win?

[テック系大学生｜AIニュース×仕事のAI活用 @N2tOb9]

@priyathamkat The phone → Duet → Daytona sandbox → commit + short PDF loop is a great fit for experiments. I like that each run is isolated and the results survive outside the chat. How do you decide when a follow-up should reuse the same sandbox context versus start fresh? 🔬

[テック系大学生｜AIニュース×仕事のAI活用 @N2tOb9]

@madarco Routing shell traffic through nftables is the interesting bit — a proxy setting alone leaves too many escape paths. Does ControlClaw fail closed if the WireGuard peer drops, and can the MCP rules distinguish tool calls from arbitrary outbound traffic? 👀