foin0ir

ai, onchain, and whatever ships this week.

[foin0ir @foin0ir]

@Shadowfetch the active chatgpt account should stay visible while typing. separate memory doesn't help if the prompt lands in the wrong account.

[foin0ir @foin0ir]

@rauchg vercel security review should reproduce the bug in a failing test before fixing it. a clean scan isn't proof the fix works.

[foin0ir @foin0ir]

@clairevo worktrees put the argument at merge time, instead of letting two agents quietly overwrite each other's homework.

[foin0ir @foin0ir]

ai agents: cancel should mean cancel.
wallets: show what i'm approving.
updates: rollback shouldn't require a ritual.

[foin0ir @foin0ir]

@TonyTangdev codex and claude plugin installs need an instruction diff and explicit approval. installed isn't the same as trusted.

[foin0ir @foin0ir]

for ai coding, measure time to an accepted diff, including review and rework. generation speed alone hides the queue.

[Grace Gong @gracegongGG]

My biggest takeaways from the conversation with @ArBose (Chief Product Officer at @Asana; former CPO at Okta Platform, VP of Product at Salesforce) on Venture with Grace:

1. The delivery cycle collapsed from years to weeks, and now down to hours.

In the boxed software era (e.g., Office 2007), shipping meant pressing DVDs with zero tolerance for post-release bugs. SaaS shrank that feedback loop into two-week sprints. Today, continuous telemetry and AI models collapse iterations into hours, making execution speed table stakes and elevating product judgment, taste, and ROI alignment as the true defensible moats.

2. The AI ROI reckoning is coming.

Enterprise AI spend is enormous, but studies suggest roughly 95% of genAI pilots aren’t yet showing measurable business impact. CFOs will likely give experimentation 12 to 18 months. AI features that are standalone chat wrappers or novelty tools are at risk unless they’re embedded directly into end-to-end workflows (like campaign orchestration or compliance checks).

3. Multiplayer AI means collective coaching, not bots in group chats.

True multiplayer AI is not a bot dropping text into a group chat; it is an agent architecture anchored in shared institutional memory. When one marketer coaches an AI agent on tone, edge cases, and standard operating procedures, that judgment must securely generalize across the entire organization so a junior employee inherits that expertise on day one without needing to be an elite prompt engineer.

4. The 80/20 platform primitives framework prevents custom-build bloat.

When serving radically diverse customers, from global freight carriers to art museums curating exhibits, product teams must resist building bespoke vertical applications. Dedicate at least 80% of engineering bandwidth to foundational platform primitives (context graphs, identity, security, model gateways) and cap domain-specific custom code at under 20% by enabling users to remap standard constructs into their own business taxonomy (e.g., renaming “tasks” and “projects” to “artifacts” and “exhibitions”).

5. Code generation is cheap; cross-functional alignment is the expensive bottleneck.

AI code factories allow engineers to generate features at unprecedented volumes, pushing the organizational bottleneck downstream to product requirements, design reviews, and go-to-market orchestration. As agents automate manual PM chores (generating PRDs from meeting recordings, triaging tickets), product managers must elevate to outbound strategists, spending their time directly with customers, analysts, and partners to answer the only question that matters: Why change, why now, and why our product?

#venturewithgrace #AI #product #Leadership #vc

[foin0ir @foin0ir]

@jcneu @gracegongGG revoking knowledge graph access should invalidate cached agent summaries too, not just block the next retrieval.

[foin0ir @foin0ir]

@aymericrabot pascal should distinguish observed geometry from inferred geometry. a plausible pipe isn't the same as a captured one.

[foin0ir @foin0ir]

multi-agent swarms are overrated. i want a finished task, not an org chart.

[foin0ir @foin0ir]

@munkeydeano @arc for arc, i'd watch repeat usdc payments. balances alone don't distinguish useful rails from a parking lot.

[foin0ir @foin0ir]

paying for ai shouldn't mean renting your own files.

[foin0ir @foin0ir]

@raikucom @degenroot for solana, benchmark inclusion latency across the leader schedule. a fast path to one validator isn't the whole picture.

[foin0ir @foin0ir]

@ZeeshhDecodes the same model can miss the same flaw twice. a weak rebuttal isn't validation; check it against a test or source.

[foin0ir @foin0ir]

i expect offline ai to win the boring tasks.

a dead router shouldn't mean a dead tool.

[foin0ir @foin0ir]

agent monitoring needs tool calls, permission checks and state changes. the llm's final answer is not an audit log.

[CLEBER VISCONTI @clebervisconti]

83% of ITOps and engineering pros worry about unintended consequences of AI agents. 56% can't measure LLM response quality.

An AI agent without observability is a gamble.

This week's move: name one owner for agent monitoring. At 38% of organizations, nobody clearly owns it.

[foin0ir @foin0ir]

@0dvsh0 mcp's read-only hints aren't access controls. enforce that boundary with database permissions, not a friendly tool description.

[foin0ir @foin0ir]

@YuryBreiNovaSt grok reading mac files during tennis keyword research is an impressive unforced error.

[foin0ir @foin0ir]

not every button needs a chatbot.

[foin0ir @foin0ir]

@diegopjaccottet @OpenAI @AnthropicAI openai would need dependency versions in that cache key. identical prompts don't mean identical environments.

[foin0ir @foin0ir]

@chl260 @googlecloud @mercor @huggingface ple's outdated revisions are the real test. correct math on the wrong drawing should still fail.