Soner Sensoy (@Ksoner) header
Soner Sensoy (@Ksoner) profile picture

Engineer, Entrepreneur. cofounder of @AppSamurai @storyly_io Interested in #appmarketing #apps #adtech #mobile #engagement #startups

London, England
Joined November 2009
We are proud to contribute to this success as @AppSamurai . Türkiye mobile ecosystem highlights🚀
3
200
What I said in July The article was called the dead spiral of learning loops. The idea is simple. Imagine a student who writes their own exams, takes them, and grades them too. Their scores go up every week. Are they learning anything? No. They are just getting better at pleasing themselves. Systems that learn from their own output die in one of two ways. Either the errors pile up, because nobody outside the system ever checks the work. Or the system stops trying new things and runs out of anything new to learn. I didn't make this up from theory. I watched it happen. My own AI note-taking system ran 500 learning cycles and 94% of what it "learned" was garbage. My automated coding system once marked eight tasks as done without writing a single line of code. It checked a status label instead of checking the actual work. Everything looked healthy. Nothing was real. My conclusion: a learning loop only works if reality gets a vote. What DeepMind built The paper is called Dream RSI. RSI means recursive self-improvement, which is a fancy way of saying: an AI that makes itself better, then uses its better self to get better again. The system has two parts. One AI invents practice problems. Whole little worlds with tasks and clear rules for winning. Think of it as a teacher writing exercises. A second AI trains inside those invented worlds. Think of it as the student. Now the loop. The teacher writes exercises. The student practices. The system looks at where the student fails. The teacher writes new exercises that attack exactly those weak spots. Repeat, thousands of times, no human involved. The clever part: the teacher improves too, not just the student. The thing deciding what to learn next is itself learning. And the big result: the student got better at real tests it had never seen. Tests the teacher never saw either. Practice in invented worlds made it better in the real one. By my July argument, this should have collapsed into confident nonsense. It didn't. Here's the thing though I read the paper carefully and started laughing. Dream RSI doesn't break my rules. It follows them. All of them. And that is exactly why it works. Remember the student who grades their own exams? DeepMind never allows that. In their system, the grading is done by simple, hard checks. The code runs or it crashes. The answer is right or it's wrong. The AI never gets to say "I think I did great" and have that count. Second, they keep a set of real tests locked away that the loop can never touch or change. That's how they know the student is actually learning, not just getting comfortable in worlds built to flatter it. Third, the teacher is forced to keep hunting for the student's weaknesses. The loop is not allowed to get cozy. Fresh difficulty, every cycle. Hard grading. Locked tests. Forced exploration. That is my July checklist, item by item. So the honest headline is not "self-improvement works, the skeptics were wrong." The honest headline is: Self-improving AI works exactly when, and only when, someone builds honest grading into it. The loop didn't escape reality. It built a machine for touching reality thousands of times a day. Where my warning still stands Now the part where I stay honest instead of declaring victory. Dream RSI works in areas where right and wrong can be checked by a machine. Code either runs or it doesn't. A math answer is either correct or it isn't. That is the easy case. But think about where most learning loops actually live. Is this writing good? Was that strategy decision smart? Did our team retro produce useful lessons? Is this AI memory worth keeping? There is no machine that can check those. A judgment has to grade them. And judgments can be fooled, flattered, and slowly bent. That is where my system failed. Not because the loop was badly built, but because "is this lesson useful" has no automatic test. So the system graded itself. And self-grading is free, while checking against reality is expensive. So here is my updated view, narrower but sharper: Where answers can be checked by a machine, self-improving AI now works, and it will improve fast. Where answers need human judgment, the dead spiral is still in charge. The problem was never the loop. The problem was always the grading. Which points at something practical. The most valuable work right now might not be building smarter AI. It might be turning fuzzy judgments into checkable tests. Every time you do that, you move one more area from the "spiral" pile to the "compounding" pile. What I'm doing about it I run a small automated build system for a side project. AI writes code, another AI reviews it, I approve what ships. After my July failures I added one rule: never trust a "done" label. Count the actual work. If a session says it finished but produced no new files, no commits, no test runs, it did not finish, full stop. That rule is just a homemade version of what DeepMind did at scale. Let reality grade the work. Never let the system grade itself. If you run any kind of learning loop, agent memory, team retros, feedback cycles, here is the whole article in one move. Find every spot where your loop grades its own work. For each one, ask: can I replace this self-grade with a check that reality performs? Where you can, do it, and the loop will genuinely improve. Where you can't, lower your trust, because that loop is one lazy quarter away from repeating itself instead of learning. Two months ago I asked whether your loops touch reality. DeepMind just showed what it looks like when they do. The spiral is still out there. It just lost the argument that it's inevitable.
The Dead Spiral of Learning Loops My AI system once marked eight tasks completed. Zero lines of code were written. A different one captured 500 learnings over two years. 23 were ever used. Every team I know is building learning loops right now: retros, postmortems, agent memory, self-improving AI. Most of these loops are already dead. They keep spinning anyway. Here is what keeping ours alive has taught me, in eight rules. 1. Status flags lie. Deltas don't. Our autonomous builder read a "done" flag and graded eight backlog items complete. Zero code, zero tests. The fix: every supervisor now counts concrete work artifacts before and after each run. No new artifacts, no completion. 2. Capture is not learning. Application is. 94 percent of our auto-captured learnings were never used in a single decision. The fix was not better capture. It was a gate: a learning only counts when it gets applied and the outcome gets recorded. 3. Every metric needs a counter-metric. When a measure becomes a target, it starts lying. Ticket-close time gets gamed by deflecting customers. Learnings-captured gets gamed by capturing garbage. Close time pairs with reopen rate. Learnings captured pairs with application rate. 4. A gate nobody uses is not a gate. Our deploy approval required explaining the deploy to another AI agent in chat. It existed on paper and failed in practice. We replaced it with one button. Approve to live: fifteen seconds. 5. Silence is a failure mode. A blocked builder sat unnoticed for 40 minutes. A finished project sat parked for nine days. Flags are not enough. If the human goes quiet for N cycles, the system must escalate or throttle itself down. 6. Grounded loops still starve. A loop that never tries new things learns more and more about less and less. James March called this the competency trap in 1991. Exploration is not the opposite of learning. It is the input of learning. 7. Damp the optimizer or it eats itself. Our growth agents run with: 20 percent of output reserved for untested ideas, a 60 percent concentration cap on any single type, weight decay with a four-week half-life, per-cycle step limits, and a weekly report that flags when one type starts eating the system. Undamped positive feedback drifts everything to one asymptote. 8. A human defines "better," on a cadence. Fast loops should never edit their own goals. Somewhere, on a schedule, a human asks whether better still means what it used to mean. The one-question audit for any loop you run: what would it look like if it were lying to me, and how would I know? A loop that only consumes its own outputs does not learn. It rehearses.
2
147
Soner Sensoy retweeted
Get ready for all of the investors in frontier model companies to suddenly flip to regulatory capture mode today Their investments in frontier model companies are currently at risk/capped because the biggest buyers of tokens are embracing open-source models: vertical AI companies, the government and enterprises That’s all this is about: stopping open source — which I’ve been saying on the pod for two years If you want safety, you want disclosure — and open source is the ultimate disclosure process
401
996
181
8,414
1,025,232
Elon is right
2
108
My favorite architecture to build something right now: owner<->operator<->builder . Owner is me, operator is the agent who has the context about my project and me and builder. Builder runs 7/24 whatever is speced by operator. Operator specs what I want and takes care of whatever needed builder cant do that or wont do like devops, marketing, ideation etc. I only work with operator and operator only work with builder. I dont even understand their language. Both are claude code. Operator lives inside my laptop, builder lives inside an aws server conputer. The communication between all 3 of us are documents and all documented and remembered.
1
285
The Dead Spiral of Learning Loops My AI system once marked eight tasks completed. Zero lines of code were written. A different one captured 500 learnings over two years. 23 were ever used. Every team I know is building learning loops right now: retros, postmortems, agent memory, self-improving AI. Most of these loops are already dead. They keep spinning anyway. Here is what keeping ours alive has taught me, in eight rules. 1. Status flags lie. Deltas don't. Our autonomous builder read a "done" flag and graded eight backlog items complete. Zero code, zero tests. The fix: every supervisor now counts concrete work artifacts before and after each run. No new artifacts, no completion. 2. Capture is not learning. Application is. 94 percent of our auto-captured learnings were never used in a single decision. The fix was not better capture. It was a gate: a learning only counts when it gets applied and the outcome gets recorded. 3. Every metric needs a counter-metric. When a measure becomes a target, it starts lying. Ticket-close time gets gamed by deflecting customers. Learnings-captured gets gamed by capturing garbage. Close time pairs with reopen rate. Learnings captured pairs with application rate. 4. A gate nobody uses is not a gate. Our deploy approval required explaining the deploy to another AI agent in chat. It existed on paper and failed in practice. We replaced it with one button. Approve to live: fifteen seconds. 5. Silence is a failure mode. A blocked builder sat unnoticed for 40 minutes. A finished project sat parked for nine days. Flags are not enough. If the human goes quiet for N cycles, the system must escalate or throttle itself down. 6. Grounded loops still starve. A loop that never tries new things learns more and more about less and less. James March called this the competency trap in 1991. Exploration is not the opposite of learning. It is the input of learning. 7. Damp the optimizer or it eats itself. Our growth agents run with: 20 percent of output reserved for untested ideas, a 60 percent concentration cap on any single type, weight decay with a four-week half-life, per-cycle step limits, and a weekly report that flags when one type starts eating the system. Undamped positive feedback drifts everything to one asymptote. 8. A human defines "better," on a cadence. Fast loops should never edit their own goals. Somewhere, on a schedule, a human asks whether better still means what it used to mean. The one-question audit for any loop you run: what would it look like if it were lying to me, and how would I know? A loop that only consumes its own outputs does not learn. It rehearses.
2
1
1
291
Dev Day, but the audience is AI agents. Announcements they can parse. APIs they can call. Docs they can read. Humans will watch the recap. My agent would definitely attend.
120
Just built this for my 6 years old as now the holiday time and she always describe games instead of telling exact names (ofcourse). nexusgaming.store/ This is an llm based game knowledge which finds you the best IOS game that you describe.
120
Developer = someone who builds what's specified. Given a problem with known shape, they produce working code. Strength: execution speed, framework fluency, pattern matching to known solutions. The job is translation, turning a spec into syntax. Engineer = someone who decides what should be built and why. Given ambiguity, they define constraints, make tradeoffs, design systems that survive change. The job is judgment under uncertainty, choosing between bad options when no good ones exist. Which one do you think will be wiped out by AI?
1
1
3
58
Companywide knowledgebase solved.
We’re introducing Workspace Intelligence. With the advanced reasoning of Gemini, we're eliminating context silos across your Workspace apps. Think of it as a unified intelligence layer that works across Workspace. See how Workspace Intelligence connects the dots and lets AI do the heavy lifting for you in this demo. goo.gle/4tkQWHq #GoogleCloudNext
4
2
81
We’re introducing Workspace Intelligence. With the advanced reasoning of Gemini, we're eliminating context silos across your Workspace apps. Think of it as a unified intelligence layer that works across Workspace. See how Workspace Intelligence connects the dots and lets AI do the heavy lifting for you in this demo. goo.gle/4tkQWHq #GoogleCloudNext
37
115
21
949
124,106
This is really cool!
Introducing workspace agents in ChatGPT—shared agents that can handle complex tasks and long-running workflows across tools and teams.
1
34
Saas is dead applies to new saas that you are vibe coding now.
43
Realized my Claude Code session can be triggered by any log line. tail -F app.log | grep X → each match becomes a prompt. The agent reads the context, decides what to do, stays quiet otherwise. No cron. No webhook. No new integration layer. Just your existing logs, reactive. Is it because Opus 4.7 very proactive?
1
56
Next updates probably will be more about claude code/cowork being more proactive just like clawd bot. Lets see...
1
98
Soner Sensoy retweeted
Next up: Liverpool FC #LIVGS #UCL
172
1,302
320
23,433
825,209
Soner Sensoy retweeted
Most founders think speed always leads to success. 10 years in: Endurance > speed. It took years to create a success story. Most would've quit by year 2. The ones still standing win. Not the ones who started fastest.
1
1
1
123
I was skipping gym recently. I downloaded claude cowork. Now it is even paying for gym. Lol
1
64
This is amazing👌
Group chats in ChatGPT are now piloting in Japan, New Zealand, South Korea, and Taiwan. A new way to collaborate with friends, family, or coworkers and ChatGPT in the same conversation.
77