Martin

Systems engineer. I post the tech thing the launch crowd skips. Often unpopular, usually unfashionable, sometimes right.

[Martin @Unpopular_Tech]

Everyone knows ChatGPT.

Everyone pays for Notion.

Everyone's heard of Grammarly.

This is for the tools that don't have a Super Bowl ad.

Every week I find something most people have never heard of that does the job better, cheaper, or cleaner than the tool everyone defaults to.

Expect: tools you've never seen on a "top 10 AI" list, honest comparisons to the expensive equivalent, who it's actually useful for and who should skip it.

If you're tired of paying for software you could replace tomorrow, follow me.

[Martin @Unpopular_Tech]

@ThisIsFormu1a1 Bearman is fine at Haas, the seat is the problem

[Martin @Unpopular_Tech]

@rohanpaul_ai formal trust models, cryptographic proofs, verifiable certificates. the frame work for giving AI a task is now more complex than the task. that's either necessary or a sign the problem is being solved at the wrong layer

[Martin @Unpopular_Tech]

@testingcatalog persistent memory in a work tool cuts both ways. less re explaining every Monday morning is real, so is the moment it confidently uses outdated context from three months ago on something that matters

[Martin @Unpopular_Tech]

@HedgieMarkets cutting seniors to fund AI, graduating juniors who can't code without it. nobody's mapped those two decisions onto the same org chart yet

[Martin @Unpopular_Tech]

@kimmonismus $16 input, $80 output, 5x spread between reading and generating, someone at Anthropic knows exactly where the cost is and it's not in the prompt

[Martin @Unpopular_Tech]

@mark_k @AnthropicAI the 52x number is a code optimization loop with a fixed goal, that's not recursive self improvement. that's a very fast auto complete on a problem someone else defined

[Martin @Unpopular_Tech]

DeepSeek built global frontier models without outside capital, got praised by Silicon Valley for capital efficiency, and is now raising $7.4B from Tencent and CATL

founder Liang Wenfeng is putting in roughly 40% himself

the "scrappy underdog" narrative and the $59B valuation are going to be awkward roommates

[Martin @Unpopular_Tech]

@49agents Velocity is easy to measure, quality is expensive to measure, that's why teams often optimize for the wrong thing

[Martin @Unpopular_Tech]

most teams using AI coding tools are not shipping better software
they're shipping more software faster with more bugs that take longer to review
the output went up, the quality did not, nobody is tracking that because the velocity number looks great in the all hands

[Martin @Unpopular_Tech]

@Cocoaisadog @heyshrutimishra you don't, at least not in the traditional sense, you test inputs and outputs, measure failure modes, and build guardrails around them

[Martin @Unpopular_Tech]

@caps_raunak

[Martin @Unpopular_Tech]

@mark_k

[Martin @Unpopular_Tech]

Dry or specific production deployment faces challenges, latency, cost spikes, format collapse. the bench mark numbers are real, the "drop it in as a replacement for your existing API calls" framing in the viral post is doing more work than the paper claims, worth reading the actual abstract before shipping it

[Martin @Unpopular_Tech]

@shiri_shh @Aiswarya_Sankar the mechanism they ranked engineers on internal leaderboards by AI tool usage, then acted surprised when usage went up, the budget didn't burn itself, the incentive structure burned it, that's not an AI problem

[Martin @Unpopular_Tech]

@TheWhizzAI @nvidia Add specificity and the first lesson usually arrives when the orchestration layer and the tool layer disagree about what the task was, not a model problem, a plumbing problem

[Martin @Unpopular_Tech]

@nvidia Engineering reality orchestration, tools, skills, secure runtime. four things that all have to work together cleanly for the agent to do anything useful. NVIDIA can sell you all four, integrating them is still your problem

[Martin @Unpopular_Tech]

Microsoft Build 2026 is selling "multi agent orchestration with no SDK required" as if the hard part of enterprise automation was ever the SDK, it wasn't, It was figuring out what to automate and who owns it when it breaks, still no update on that

[Satya Nadella @satyanadella]

Building a frontier intelligence ecosystem together.

Highlights from my keynote at Microsoft Build this morning.

[Martin @Unpopular_Tech]

@tomwarren "1.4x performance per watt" is a carefully chosen metric, not faster, not cheaper, more efficient. worth knowing what work load that number came from before treating it as a general claim about the chip