Tool Teardowns
Hands-on reviews of AI tools — what works, what is a wrapper with a waitlist. Where Reachium shows up in context.
I Read the Terms of 13 AI Note-Takers. Nine Make Consent Your Problem. Zero Indemnify You.
Nine of 13 meeting-AI vendors contractually assign recording consent to the customer, and not one of the 13 indemnifies that customer against a privacy or wiretap claim.
I Read Six AI Leaderboards on One Day. They Named Six Different Number Ones.
On 2026-08-15 I pulled the top entry off seven public leaderboards. The six I could read crowned six different entries from three different base models, and every board that publishes an error bar has its number-one-versus-number-two gap sitting inside it.
I Traced 17 Humanity's Last Exam Scores. Two of Them Are on a Leaderboard That Ran the Test.
Every one of the 7 vendor-published HLE scores I checked states whether tools were switched on. Zero of the 10 scores quoted by leaderboard aggregators do, and one site's top three numbers appear on no board at all.
I Audited 200 Servers in the Official MCP Registry. One in Six Repository Links Is Dead.
Of 163 sampled MCP servers that list a source repository, 26 point at a 404, and only 45 of 200 publish a security policy at all.
15 AI Extensions, 59.5 Million Installs. The Chrome Web Store Published the Permissions for Exactly Zero of Them.
Install counts, update dates and data-collection declarations are all auditable from a Chrome Web Store listing. The one field that tells you what the extension can reach is not in the page at all.
I Read the LICENSE File in 35 'Open Source' AI Agents. Four Are Not Open Source, and n8n Is Not One of Them.
31 of 35 agent repos ship an OSI-approved license and 4 do not. Three of those four call themselves open source in their own README. n8n, the project everyone points at, calls itself fair-code.
I Ran a Five-Signal Staleness Check on 30 AI Tools I Could Still Pay For. Three Failed It.
Four of thirty showed two or more stale signals while still selling; one of the four survived the confirmation step. The worst gap was 194 days between the last shipped code and a live $500 plan.
An Agent Deleted a Production Database in 9 Seconds. I Read 20 Agent Products' Docs. Three Mention Credential Scoping.
Of 20 agent and coding-agent products, 10 document a destructive-action gate you can enforce, 6 document any rollback after a tool call has run, and 3 tell you to scope the credential the agent uses. Only one documents both a gate and a rollback.
I DNS-Audited 30 Cold-Email Vendors. Six Run p=none, Which Google Allows. One Publishes Three SPF Records, Which RFC 7208 Does Not.
SPF and DMARC lookups on the corporate domains of 30 outreach, AI-SDR and warmup vendors: all 30 publish both records, 6 sit at p=none (which Google's own rules permit), 6 have no usable reporting address, and smartlead.ai publishes three conflicting SPF records while its own help center says that breaks SPF.
I Read 25 AI-Visibility Vendors' Public Pages. Five Say Whether They Query ChatGPT or an API.
Every one of them sells a share-of-voice number. Only 5 of 25 disclose the single choice that decides what that number means, and 4 disclose none of the six things a buyer needs.
15 AI Vendors Publish an Uptime Number. I Checked Every One Against Its Own Incident Log. 11 Don't Add Up.
I audited 25 AI vendor status pages. Fifteen publish a numeric uptime percentage, and eleven of those publish a figure their own incident log cannot account for. The four that reconcile all measure uptime with a robot instead of declaring it.
I Checked 36 Named RSL Supporters for the License File. Three Have One.
RSL's own press release claims 1,500+ endorsers. Of the 36 organizations it names, three serve the license file the spec defines, 31 return a clean 404, and two are unreadable behind a 403.
I Looked for a Support Email at 25 AI Companies. Five Have One.
Before you pay, 5 of 25 AI vendors publish a support email address, 20 route you to a form or a chatbot, 11 name an AI agent as the front line, and exactly 1 publishes a response time in hours.
ChatGPT Work Is a Month Old and Almost Nobody Has Actually Tested It
OpenAI retired agent mode and shipped an agent that 'turns a goal into finished work.' The early evidence says: real for bounded, file-shaped tasks, and the review economy hasn't caught up enough to tell you more.
I Read Every Public Security Disclosure About Perplexity's Comet Browser. There Are Six.
The agentic browser is genuinely useful, free, and has racked up six publicly documented security findings in six and a half months, including a zero-click calendar-invite attack that took two patches to kill.
I Did the Math on Clay's Credits So You Don't Have To
Clay's free tier buys you exactly 200 found emails a month, its G2 and Trustpilot scores are 4.9 and 2.4 for the same product, and the viral credit-burn stats are unsourced. Here is what the official numbers actually say.
I Priced Every $100+ AI Subscription. The Full Stack Is Now $1,400 a Month.
Ten premium tiers across seven vendors, four of them selling the identical 5x-for-$100, 20x-for-$200 ladder. Whether any of them is worth it depends on one number vendors rarely say out loud.
Granola vs Fathom: The Bot Just Left the Meeting
Fathom's April update copied Granola's signature bot-free capture, and two-thirds of the comparison content written since still describes a product that no longer exists.
I Checked 30 Sites for llms.txt. The AI Companies Don't Have One.
97% of llms.txt files never get read, no AI vendor documents using them, and the labs themselves don't serve one on their main domains. The file is decoration, with one narrow exception.
Cursor vs Claude Code in 2026: The Comparison Everyone Writes and Nobody Can Price
Adoption is tied at 18% each, both vendors now call their product an agent, and their pricing units are deliberately incomparable. Here is how to actually choose.
I Audited Eight 'AI Outreach' Tools Against Their Own Websites. Here's the Body Count.
Eight vendors, their own homepages and docs, one question: where are the receipts? Several are still a Chrome extension in a trench coat, and exactly zero publish a real benchmark where buyers look.
AI Note-Takers Sell You Action Items. Not One of Eight Vendors Publishes an Accuracy Number for Them.
The transcription is close to solved. The 'action items' are unaudited. And the consent machinery all eight vendors built tells you what a bot in the room does to a conversation.
Built by the team that actually does outbound.
AI Until I Die is written by the operators behind Reachium — the platform that runs the LinkedIn outreach we write about. Same brain, live product, real receipts.
See what Reachium does ↗