Marc Bara

Project Management, Technology, Education. PMP-certified Project Manager. PhD on Signal Processing.
Founder of

[Marc Bara @marc_bara]

@AnthropicAI 9/
The 8x headline is real. So is the 4x reality. The distance between them is a law of systems, not a failure of ambition.

Read the full analysis here: https://t.co/P162D9ljBN

[Marc Bara @marc_bara]

1/
@AnthropicAI just published "When AI Builds Itself." More than 80% of code merged into their codebase is now written by Claude. Engineers are shipping 8x as much code per day as in 2024. On one benchmark, Claude hit a 52x speedup on training code optimization — a skilled human researcher takes 4 to 8 hours to reach 4x. Those numbers are real. But they describe only part of the system. And that distinction is everything. 🧵

[Marc Bara @marc_bara]

8/
The irony worth sitting with: this is a luxury problem. Anthropic hit the review and judgment ceiling because they actually automated execution at scale. A May 2026 survey by Writer found that 79% of organizations are still struggling to turn AI adoption into measurable business value, and 38% of small and mid-sized companies have not automated a single workflow yet. Amdahl's Law does not kick in until you have actually accelerated something. Most organizations are not there yet.

[Marc Bara @marc_bara]

7/
This also changes how to read the next few releases. The question is no longer whether a feature can do something new. It is how much it removes from doing something you could already do. The releases that matter now are the ones that take work away.

Read the full analysis here:
https://t.co/mBUsDhEMIX

[Marc Bara @marc_bara]

1/
On May 28, @AnthropicAI added dynamic workflows to Claude Code. The pitch: give it a plain-language request, it writes its own orchestration script, spins up hundreds of subagents in parallel, checks their work, and returns a single verified result. That sounds like a leap. Strip it to its parts and the leap mostly disappears. 🧵

[Marc Bara @marc_bara]

6/
The practical implication is direct. The skill that separated people who could marshal a fleet of agents from those who could not, the ability to design and run the orchestration by hand, is moving from an advantage to a baseline. If that was your edge, assume it depreciates. The edge moves up the stack: to judgment about which problems are worth a fleet, and to knowing how to verify what one hands back.

[Marc Bara @marc_bara]

7/
Before you celebrate any AI productivity figure, find the exact stage it accelerated, then trace that stage to the output that actually leaves the building. If the two numbers diverge the way they do in this study, the smaller one is your result. That gap is the diagnostic — it points to the human-bound step that has become the binding constraint, and that is worth more than the 17x ever was.

Read the full analysis here:
https://t.co/nBSjWgA05o

[Marc Bara @marc_bara]

1/
AI tools write 17x more lines of code. The same work produces only 1.3x more software releases. A study of 100,000 @GitHub developers just put exact numbers on where all that productivity goes. 🧵

[Marc Bara @marc_bara]

6/
Lines of code is an activity metric at the start of a chain. Releases is a throughput metric at the end of it. Activity metrics sit where the tool is and show the biggest number. Track only that, and you can report a 17x that never reaches the output the business actually counts. The pattern generalizes: a contract drafter that writes ten times faster does not close ten times more deals if legal review and client sign-off are unchanged. The gain you can capture is capped by the slowest human-bound stage downstream of the one you accelerated.