Emre Coklar https://emrecoklar.com Writing, dated bets, and tutorials from someone who uses AI as the execution layer — then shows what changed because of it. en-us I Retired This Project in July. Its Daemon Is Still Running. https://emrecoklar.com/writing/retired-in-july-still-running https://emrecoklar.com/writing/retired-in-july-still-running Sun, 23 Aug 2026 00:00:00 GMT Documentation records intentions; the machine records reality. Four probes that found four contradictions in my own AI workspace, and the one drift no command could catch. RLM removes the context ceiling. The unlock is at the intersection of RLM, retrieval that stuffs the REPL with existing art, and memory that distinguishes what is forever-true from what should be forgotten. https://emrecoklar.com/bets/rlm-retrieval-memory https://emrecoklar.com/bets/rlm-retrieval-memory Sat, 15 Aug 2026 00:00:00 GMT Context windows were never going to be the architecture. Once a model can operate over more than it can hold, the question becomes what you put in the environment and what you refuse to forget. The right harness shape is RLM + GEPA + dspy.FLEX, with expensive frontier models driving cheaper models as workhorses. https://emrecoklar.com/bets/rlm-gepa-dspy-flex-cascade https://emrecoklar.com/bets/rlm-gepa-dspy-flex-cascade Sat, 15 Aug 2026 00:00:00 GMT I am logging this instead of building it. Nobody has monetized a harness, because the labs give them away with API access. Building it would only be something I was proud of. The agent harness is solved. Synthesis is not. https://emrecoklar.com/bets/harness-is-solved-synthesis-isnt https://emrecoklar.com/bets/harness-is-solved-synthesis-isnt Wed, 05 Aug 2026 00:00:00 GMT Prime Intellect shipped Prime Agent, MIT-licensed, above the human-expert baseline on ARC-AGI-3. The RLM runtime is about 1,500 lines. Decomposition is a public result. Personification with In-App Characters https://emrecoklar.com/writing/personification-with-in-app-characters https://emrecoklar.com/writing/personification-with-in-app-characters Sat, 01 Aug 2026 00:00:00 GMT A gated, agent-operated research sprint that took Celstate from paper cutouts to a character persona worth shipping, told through the checkpoint artifacts left along the way. Long context windows buy retrieval. They do not buy correlation. https://emrecoklar.com/bets/long-context-buys-retrieval-not-correlation https://emrecoklar.com/bets/long-context-buys-retrieval-not-correlation Thu, 06 Feb 2025 00:00:00 GMT Posted this on 2025-02-06, when the industry was celebrating million-token windows as if stuffing more text into the prompt solved reasoning across it.