Writing by Alnur Ismail on startups, AI, and the intersection between them.https://www.alnurismail.com/enCopyright © 2026 aiaiaiFri, 15 May 2026 20:26:09 GMTAstro RSShttps://www.rssboard.org/rss-specification60https://www.alnurismail.com/posts/llm-as-a-judge/https://www.alnurismail.com/posts/llm-as-a-judge/Notes on LLM-as-a-judge, calibration, overconfidence, and why evaluation confidence is harder than it looksFri, 15 May 2026 00:00:00 GMTLLM-as-a-Judgehttps://www.alnurismail.com/notes/llm-as-a-judge/image.webpLLM-as-a-JudgeEvaluationPaperClubaihttps://www.alnurismail.com/posts/ai-agent-traps-into-product-tests/https://www.alnurismail.com/posts/ai-agent-traps-into-product-tests/Open source framework designed to test how AI products behave when adversarial content enters their normal workflowMon, 11 May 2026 00:00:00 GMTsecurityhttps://www.alnurismail.com/notes/ai-agent-traps-into-product-tests/adobe_am.webpAgentsSecurityaihttps://www.alnurismail.com/posts/ai-agent-poison/https://www.alnurismail.com/posts/ai-agent-poison/Manipulating, deceiving, and exploiting agentsWed, 08 Apr 2026 00:00:00 GMTsecurityhttps://www.alnurismail.com/notes/ai-agent-poison/image.webpAgentsSecurityaihttps://www.alnurismail.com/posts/experience-reflect-consolidate/https://www.alnurismail.com/posts/experience-reflect-consolidate/Read a few papers around the reflection loop.Mon, 16 Mar 2026 00:00:00 GMTCompilationhttps://www.alnurismail.com/notes/experience-reflect-consolidate/image.webpAgenticCompilationPaperClubaihttps://www.alnurismail.com/posts/action-dispatch-execute/https://www.alnurismail.com/posts/action-dispatch-execute/Realizing dispatch is not optional in tool calling systemsThu, 19 Feb 2026 00:00:00 GMTToolshttps://www.alnurismail.com/notes/action-dispatch-execute/image.webpAgenticToolsaihttps://www.alnurismail.com/posts/mystro-sandboxing-tooling/https://www.alnurismail.com/posts/mystro-sandboxing-tooling/Moving execution into a sandbox and rethinking tool interaction and validationTue, 10 Feb 2026 00:00:00 GMTToolshttps://www.alnurismail.com/notes/mystro-sandboxing-tooling/image.webpAgenticToolsaihttps://www.alnurismail.com/posts/mystro-gepa-run/https://www.alnurismail.com/posts/mystro-gepa-run/Integrating GEPA into the agentic loopThu, 05 Feb 2026 00:00:00 GMTGEPAhttps://www.alnurismail.com/notes/mystro-gepa-run/image.webpGEPAAgenticaihttps://www.alnurismail.com/posts/prompt-optimization/https://www.alnurismail.com/posts/prompt-optimization/Building a system for evolutionary prompt optimization in agentic codingWed, 28 Jan 2026 00:00:00 GMTGEPAhttps://www.alnurismail.com/notes/prompt-optimization/image.webpGEPAAgenticaihttps://www.alnurismail.com/posts/vectors-not-tokens-for-bci/https://www.alnurismail.com/posts/vectors-not-tokens-for-bci/Exploring non-invasive BCI that goes beyond backpropFri, 16 Jan 2026 00:00:00 GMTbci, spahttps://www.alnurismail.com/notes/vectors-not-tokens-for-bci/image.webpBCISPAaihttps://www.alnurismail.com/posts/ablation-jepa-standalone/https://www.alnurismail.com/posts/ablation-jepa-standalone/Went deep down the rabit hole to validate what parts of the model work and don'tFri, 12 Dec 2025 00:00:00 GMThttps://www.alnurismail.com/notes/ablation-jepa-standalone/image.webpRLJEPAaihttps://www.alnurismail.com/posts/validating-jepa-standalone/https://www.alnurismail.com/posts/validating-jepa-standalone/Trained a JEPA model and shifted focus to validating itTue, 09 Dec 2025 00:00:00 GMThttps://www.alnurismail.com/notes/validating-jepa-standalone/image.webpRLJEPAaihttps://www.alnurismail.com/posts/jepa-valence-policy-standalone/https://www.alnurismail.com/posts/jepa-valence-policy-standalone/Designing an experiment to validate if JEPA can actually learn a representationThu, 04 Dec 2025 00:00:00 GMThttps://www.alnurismail.com/notes/jepa-valence-policy-standalone/image.webpRLJEPAaihttps://www.alnurismail.com/posts/jepa-valence-policy/https://www.alnurismail.com/posts/jepa-valence-policy/Created a policy using a JEPA world model that doesn't work very wellMon, 01 Dec 2025 00:00:00 GMThttps://www.alnurismail.com/notes/jepa-valence-policy/image.webpRLJEPAaihttps://www.alnurismail.com/posts/paper-club-curiosity/https://www.alnurismail.com/posts/paper-club-curiosity/Read a few papers on intrinsic motivationWed, 26 Nov 2025 00:00:00 GMTcuriosityhttps://www.alnurismail.com/notes/paper-club-curiosity/image.webpCuriosityPaperClubRLaihttps://www.alnurismail.com/posts/stochastic-policy-cma-es/https://www.alnurismail.com/posts/stochastic-policy-cma-es/Changing the policy to use probabilities solved the observed action space collapseMon, 24 Nov 2025 00:00:00 GMTcma-eshttps://www.alnurismail.com/notes/stochastic-policy-cma-es/image.webpCMA-ESESRLaihttps://www.alnurismail.com/posts/reward-shaping-bitter-lesson/https://www.alnurismail.com/posts/reward-shaping-bitter-lesson/Introducing a more complex action space lead to bitter lessonThu, 20 Nov 2025 00:00:00 GMTcma-eshttps://www.alnurismail.com/notes/reward-shaping-bitter-lesson/heatmap.webpCMA-ESESRLaihttps://www.alnurismail.com/posts/multilayer-cma/https://www.alnurismail.com/posts/multilayer-cma/Understanding how to structure a multilayer CMA-ES networkFri, 14 Nov 2025 00:00:00 GMTcma-eshttps://www.alnurismail.com/notes/multilayer-cma/mlp_cmaes.webpCMA-ESESaihttps://www.alnurismail.com/posts/cmaes-policy/https://www.alnurismail.com/posts/cmaes-policy/Understanding the structure of a CMA-ES policy and training loop implementationFri, 31 Oct 2025 00:00:00 GMTcma-es policyhttps://www.alnurismail.com/notes/cmaes-policy/image.webpCMA-ESRLESaihttps://www.alnurismail.com/posts/cma-es-for-agent-training/https://www.alnurismail.com/posts/cma-es-for-agent-training/Applying CMA-ES to optimize agent behavior through parameter searchTue, 28 Oct 2025 00:00:00 GMTcma es policy search agent traininghttps://www.alnurismail.com/notes/cma-es-for-agent-training/image.webpCMA-ESRLESaihttps://www.alnurismail.com/posts/cma-es-cov-ellipse/https://www.alnurismail.com/posts/cma-es-cov-ellipse/Visualizing CMA-ES search behavior and understanding step sizeMon, 27 Oct 2025 00:00:00 GMTcma es scatter plot optimization ellipsehttps://www.alnurismail.com/notes/cma-es-cov-ellipse/image.webpCMA-ESESaihttps://www.alnurismail.com/posts/covariance-matrix-and-distribution/https://www.alnurismail.com/posts/covariance-matrix-and-distribution/Understanding CMA-ES, NES, and multivariate Gaussian conceptsFri, 24 Oct 2025 00:00:00 GMTcovariance multivariate gaussian evolution strategieshttps://www.alnurismail.com/notes/attachments/cma-es.webpCMA-ESESaihttps://www.alnurismail.com/posts/learning-without-vision/https://www.alnurismail.com/posts/learning-without-vision/Designing curiosity loops for simple survival-based agentsWed, 22 Oct 2025 00:00:00 GMTcuriosityhttps://www.alnurismail.com/notes/attachments/cma-es.webpcuriositymotivationaihttps://www.alnurismail.com/posts/evo-strats/https://www.alnurismail.com/posts/evo-strats/Exploring neuroevolution, curiosity, and the biological roots of learningTue, 21 Oct 2025 00:00:00 GMTevolutionary strategieshttps://www.alnurismail.com/notes/evo-strats/image.webpcuriosityevolutionary-strategiesaihttps://www.alnurismail.com/posts/causal-world-models-overview/https://www.alnurismail.com/posts/causal-world-models-overview/Exploring causal reasoning and world models in AIMon, 08 Sep 2025 00:00:00 GMTcausal reasoning bengio gflownets world modelshttps://www.alnurismail.com/notes/attachments/world_model.webpcausal-mlworld-modelsaihttps://www.alnurismail.com/posts/latent-space-in-ai/https://www.alnurismail.com/posts/latent-space-in-ai/Thinking through what latent space means and how it connects to representation learningWed, 23 Jul 2025 00:00:00 GMTlatent space representation alberta plan featureshttps://www.alnurismail.com/notes/attachments/latent-space.webplatent-spacerepresentation-learningaihttps://www.alnurismail.com/posts/text-similarity-approaches/https://www.alnurismail.com/posts/text-similarity-approaches/Exploring ASR model confidence and similarity scoring methodsWed, 11 Jun 2025 00:00:00 GMTasr similarity phonemes confidencehttps://www.alnurismail.com/notes/text-similarity-approaches/image.webpASRtext-similarityaihttps://www.alnurismail.com/posts/understanding-latent-space/https://www.alnurismail.com/posts/understanding-latent-space/Exploring methods to visualize and analyze latent spaces in VAEs and JEPATue, 10 Jun 2025 00:00:00 GMTlatent visualization vae jepa umap tsnehttps://www.alnurismail.com/notes/attachments/latent-space.webplatent-spaceaihttps://www.alnurismail.com/posts/encoder-neural-network-explanation/https://www.alnurismail.com/posts/encoder-neural-network-explanation/Clarifying how encoders work and how to structure them for JEPAFri, 21 Feb 2025 00:00:00 GMTencoder neural network compression jepahttps://www.alnurismail.com/notes/encoder-neural-network-explanation/image.webpencodersJEPAaihttps://www.alnurismail.com/posts/masking-strategies-for-jepa/https://www.alnurismail.com/posts/masking-strategies-for-jepa/Understanding deterministic masking and patching for JEPA datasetsWed, 19 Feb 2025 00:00:00 GMTjepa masking deterministic seed lightninghttps://www.alnurismail.com/notes/attachments/jepa-mask.webpJEPAmaskingaihttps://www.alnurismail.com/posts/jepa-gta5-world-generation/https://www.alnurismail.com/posts/jepa-gta5-world-generation/Applying JEPA to generate GTA5-like worlds from frame dataTue, 18 Feb 2025 00:00:00 GMTjepa world model gta5 frame predictionhttps://www.alnurismail.com/notes/attachments/jepa-mask.webpJEPAworld-modelsaihttps://www.alnurismail.com/posts/normalization-in-ml-explained/https://www.alnurismail.com/posts/normalization-in-ml-explained/Understanding integration, PDFs, and normalization in MLFri, 31 Jan 2025 00:00:00 GMTnormalization pdf probability integrationprobabilitynormalizationaihttps://www.alnurismail.com/posts/disentangled-vae-resources/https://www.alnurismail.com/posts/disentangled-vae-resources/Finding research papers and resources for improving disentanglement in VAEsTue, 28 Jan 2025 00:00:00 GMTbeta vae disentanglement representation learningVAEdisentanglementaihttps://www.alnurismail.com/posts/dynamic-vae-layer-creation/https://www.alnurismail.com/posts/dynamic-vae-layer-creation/Building encoder and decoder layers dynamically based on input parametersWed, 22 Jan 2025 00:00:00 GMTpytorch dynamic layers encoder decoder vaeVAEaihttps://www.alnurismail.com/posts/latent-space-rotation-vae/https://www.alnurismail.com/posts/latent-space-rotation-vae/Exploring how to navigate or rotate within a VAE latent spaceWed, 22 Jan 2025 00:00:00 GMTlatent vector manipulation 600d vaeVAElatent-spaceaihttps://www.alnurismail.com/posts/high-val-loss-diagnosis/https://www.alnurismail.com/posts/high-val-loss-diagnosis/Understanding why validation loss is high when training a VAE and what KLD really meansThu, 16 Jan 2025 00:00:00 GMTkl divergence reconstruction loss vae pytorchVAEKLDloss-analysisaihttps://www.alnurismail.com/posts/vae-rgb-importance/https://www.alnurismail.com/posts/vae-rgb-importance/Debugging color and visualization issues when training VAEs on RGB dataThu, 16 Jan 2025 00:00:00 GMTrgb tensorboard pytorch lightning vaeVAECVaihttps://www.alnurismail.com/posts/learning-vaes-and-projects/https://www.alnurismail.com/posts/learning-vaes-and-projects/Starting to learn VAEs and figuring out how to apply them to world generationTue, 14 Jan 2025 00:00:00 GMTvariational autoencoder gta5 procedural world generationVAEworld-modelsaihttps://www.alnurismail.com/posts/approximate-inference-simplified/https://www.alnurismail.com/posts/approximate-inference-simplified/Breaking down probabilistic models and VAEs to understand inference, ELBO, and KL divergenceTue, 07 Jan 2025 00:00:00 GMTinference vae kl elbo probabilisticVAEKLDaihttps://www.alnurismail.com/posts/learning-about-emergent/https://www.alnurismail.com/posts/learning-about-emergent/Understanding exponential forms and attention variants in transformersThu, 02 Jan 2025 00:00:00 GMTexponential loss grouped query rotary embeddingsattentiontransformersaihttps://www.alnurismail.com/posts/probabilistic-programming-learning-plan/https://www.alnurismail.com/posts/probabilistic-programming-learning-plan/Building a learning plan for probabilistic programming in AIMon, 09 Dec 2024 00:00:00 GMTprobabilistic programming bayes net vikash mansinghkaprobabilistic-programmingresearchaihttps://www.alnurismail.com/posts/novel-ai-approaches/https://www.alnurismail.com/posts/novel-ai-approaches/Looking at new directions in AI that go beyond deep neural networksTue, 26 Nov 2024 00:00:00 GMTevolutionary computation neips research velocityresearchaihttps://www.alnurismail.com/posts/understanding-state-space-models/https://www.alnurismail.com/posts/understanding-state-space-models/Breaking down Mamba’s state equations and how they relate to ML model outputsTue, 29 Oct 2024 00:00:00 GMTmamba ssm hprime state-equationmambaSSMaihttps://www.alnurismail.com/posts/wav2vec2-has-12-layers/https://www.alnurismail.com/posts/wav2vec2-has-12-layers/Inspecting transformer layers and hidden states in Wav2Vec2Thu, 04 Jul 2024 00:00:00 GMTwav2vec2 transformer hidden statesASRwav2vecaihttps://www.alnurismail.com/posts/more-epochs-better-inference/https://www.alnurismail.com/posts/more-epochs-better-inference/Understanding the tradeoffs between epochs, data, and overfittingWed, 19 Jun 2024 00:00:00 GMTepochs overfitting training generalizationtraininggeneralizationaihttps://www.alnurismail.com/posts/agents-thinking/https://www.alnurismail.com/posts/agents-thinking/Connecting agentic workflows to tree-of-thought and reinforcement learning ideasThu, 13 Jun 2024 00:00:00 GMTagentic workflows tree of thought reinforcement learningagentic-systemsreasoningaihttps://www.alnurismail.com/posts/speechbrain-structure/https://www.alnurismail.com/posts/speechbrain-structure/Exploring SpeechBrain's modular system of lobes and recipes for ASR and representation learning.Thu, 13 Jun 2024 00:00:00 GMTspeechbrainhttps://www.alnurismail.com/notes/speechbrain-structure/speechbrain_modules.webpspeechbrainASRaihttps://www.alnurismail.com/posts/audio-feature-inter/https://www.alnurismail.com/posts/audio-feature-inter/Investigating how to reconstruct or understand the 768-dim latent features from wav2vec 2.0.Thu, 16 May 2024 00:00:00 GMTasrhttps://www.alnurismail.com/notes/audio-feature-inter/latent_features.webpASRlatent-spacewav2vecaihttps://www.alnurismail.com/posts/rep-learning-in-wav2vec/https://www.alnurismail.com/posts/rep-learning-in-wav2vec/Exploring how wav2vec 2.0 and vq‑wav2vec learn speech representations via self‑supervised objectives.Thu, 16 May 2024 00:00:00 GMTwav2vec self supervised speech representationhttps://www.alnurismail.com/notes/rep-learning-in-wav2vec/wav2vec_diagram.webpwav2vecASRaihttps://www.alnurismail.com/posts/phoneme-models/https://www.alnurismail.com/posts/phoneme-models/Evaluating SpeechBrain's support for phoneme-level ASR and model discovery challenges.Thu, 02 May 2024 00:00:00 GMTasrhttps://www.alnurismail.com/notes/phoneme-models/missing_model.webpASRphoneme-recognitionai