Deep End Era. Engineering with Sankalp

Sankalp Thakur

Long-form engineering, computing and mathematics. Created by Sankalp Thakur. A Transient Labs production. Website: https://transientlabs.ai/the-deep-end GitHub: https://github.com/sankalpsthakur X: https://x.com/sankalpsthakur LinkedIn: https://www.linkedin.com/in/sankalpthakur/ Hugging Face: https://huggingface.co/sankalpsthakur Scripts: Claude Opus. Synthetic narration: Edge TTS. Reviewed, listened to and approved by Sankalp Thakur.

  1. Oct 12

    Running the Heat Equation Backwards: Diffusion Models and Flow Matching

    Pipeline integrity engineer Wren Okafor watches a dye pulse smear out in a water line, and her friend Teodor Lindqvist, a machine learning engineer, claims his whole field runs that smearing backwards. Starting from the score (the gradient of log density, which makes the normalising constant vanish), the episode builds Langevin dynamics, Song and Ermon's annealed noise-conditional score networks, and Ho, Jain and Abbeel's DDPM (2020): the variance-preserving forward process, the noise-prediction loss, and its equivalence to denoising score matching. It then covers the score-SDE view (Song et al., ICLR 2021), Anderson's reverse-time SDE, the Fokker–Planck equation (advection plus diffusion, and literally the heat equation in the variance-exploding case), and the deterministic probability-flow ODE. Next come fast samplers (DDIM, DPM-Solver, Karras's EDM), classifier and classifier-free guidance, latent diffusion and Stable Diffusion, and diffusion transformers. The second half moves from noise to transport: continuous normalising flows and the continuity equation, flow matching (Lipman et al.), rectified flow and reflow (Liu et al.), stochastic interpolants (Albergo and Vanden-Eijnden), Monge, Kantorovich and Benamou–Brenier optimal transport and minibatch OT, then consistency models and distillation. It closes with the state of the art, each claim tiered: SD3's rectified-flow MM-DiT, FLUX.1, FLUX.2 and FLUX 3, video models, masked diffusion language models (LLaDA, Mercury, Gemini Diffusion), Diffusion Policy and pi-zero in robotics, RFdiffusion and AlphaFold 3 in protein science, and GenCast in weather. Builds on ep69 (PDEs), ep70 (functional analysis), ep15 (information theory), ep22 (probability) and ep34 (inference economics). Chapters 00:00:00 Cold open: the ink drop in the pipe 00:03:13 The problem: drawing from a distribution you can't write down 00:06:04 The score and why the constant disappears 00:09:18 Langevin dynamics: rolling downhill with a shaker attached 00:12:34 DDPM: the forward process, a slow fade to static 00:15:46 DDPM: the reverse process and the beautifully simple loss 00:19:04 The score-SDE view: let the steps become infinitely small 00:22:02 Fokker–Planck: from particles to the density 00:24:40 The probability-flow ODE: removing the dice 00:27:32 DDIM and the race to fewer steps 00:30:51 Guidance: steering the sample toward a prompt 00:34:06 Latent diffusion and Stable Diffusion 00:37:00 Continuous normalising flows: transport without noise 00:40:17 Flow matching: regress the velocity 00:43:16 Rectified flow: make the paths straight 00:45:27 Stochastic interpolants: the unifying frame 00:47:44 Optimal transport: the shortest way to move the sand 00:50:00 Consistency models and distillation: from fifty steps to one 00:53:20 The state of the art, part one: images and video 00:56:59 The state of the art, part two: diffusion for language 00:59:57 The state of the art, part three: robots, proteins and weather 01:03:38 What it means, and what to watch Created by Sankalp Thakur. A Transient Labs production. Website: https://transientlabs.ai GitHub: https://github.com/sankalpsthakur X / Twitter: https://x.com/sankalpsthakur LinkedIn: https://www.linkedin.com/in/sankalpthakur/ Hugging Face: https://huggingface.co/sankalpsthakur Scripts generated with Claude Opus. Synthetic narration generated with Edge TTS. Reviewed, listened to, and approved by Sankalp Thakur.

    Running the Heat Equation Backwards: Diffusion Models and Flow Matching

  2. Oct 12

    Planning, Verification and Building Your Own (Harness Design, Part 3)

    Ines's agent now either gives up too early or keeps going after it's done. Fan, Zhang et al.'s planning ablation explains both. A persistent to-do list added 11.6 points for Nemotron-3 30B (median run 5 turns without it, 40 with it). It cut SWE-Bench cost by about 30% for Nemotron-3 550B and 32% for Mistral Medium 3.5 by trimming redundant verification, with a small, non-significant drop in accuracy. The episode covers verification and tests that give wrong answers, SWE-Bench Verified contamination and the SWE-Bench Pro retraction, the harness as a hidden variable in leaderboards, latency and cost, critiques of the paper, how to run your own ablation, a 13-point design checklist, and how it all maps onto multi-agent setups. Chapters 00:00:00 Cold open: what the agent does with more time 00:02:03 What planning means in this paper 00:04:23 The weak model: a list that keeps it alive 00:06:58 The strong models: a list that tells them when to stop 00:09:52 What a plan looks like in practice 00:12:40 The middle model and the shape of a distribution 00:14:45 Verification: the oracle and the over-checker 00:17:50 When the oracle lies 00:21:54 Evaluation pitfalls: the benchmark under the microscope 00:27:06 The harness is a confound in every leaderboard 00:29:52 Latency, turns and the shape of the bill 00:32:40 What the critics say 00:36:27 The big idea: harness components age 00:38:42 Running your own ablation 00:41:24 A checklist for building your own 00:45:22 From one agent to many 00:48:10 Closing: back at the whiteboard Created by Sankalp Thakur. A Transient Labs production. Website: https://transientlabs.ai GitHub: https://github.com/sankalpsthakur X / Twitter: https://x.com/sankalpsthakur LinkedIn: https://www.linkedin.com/in/sankalpthakur/ Hugging Face: https://huggingface.co/sankalpsthakur Scripts generated with Claude Opus. Synthetic narration generated with Edge TTS. Reviewed, listened to, and approved by Sankalp Thakur.

    Planning, Verification and Building Your Own (Harness Design, Part 3)

  3. Oct 12

    Context Is the Bottleneck (Harness Design, Part 2)

    Ines finds that more than half her agent's failed runs never wrote a line: they overflowed the context window. With Gideon she works through why the history fills up, context rot, lossy versus lossless compaction, and the five tiers in Fan, Zhang et al. (arXiv:2609.20804). On SWE-Bench Verified, managing context was worth 35.7 points at a 32k window but only 2.7 at 128k, and almost all of that came from stopping overflow (78.7% of unmanaged runs overflowed at 32k, versus zero for managed runs). Eliding first and summarising later was the cheapest policy. Recoverable recall was rarely called and added no accuracy. The episode also covers window arithmetic, surprises in the tables, how compaction fails, keeping the plan outside the history, learned versus static policies, vendor features, sub-agents and the limits of the evidence. Chapters 00:00:00 Cold open: the runs that never wrote a line 00:02:28 Why the bucket fills 00:04:54 A bigger window is not a free fix 00:07:23 Lossy and lossless: two families of fix 00:09:30 The five tiers 00:10:55 How the staged policy works 00:13:17 The arithmetic of a window 00:16:44 Finding one: management matters most when the window binds 00:19:08 The overflow story 00:22:18 Reading the tables: the surprises 00:25:56 Finding two: cheap first, expensive later 00:29:01 Finding three: the photos in the drawer 00:33:24 What the summary says 00:35:07 How compaction goes wrong 00:38:11 Why the plan lives outside the history 00:39:24 The wider toolkit: maps, memory and sub-agents 00:42:53 Static rules and learned habits 00:45:17 The vendors' versions 00:47:32 What the paper does not tell you about context 00:49:48 What Ines changed Created by Sankalp Thakur. A Transient Labs production. Website: https://transientlabs.ai GitHub: https://github.com/sankalpsthakur X / Twitter: https://x.com/sankalpsthakur LinkedIn: https://www.linkedin.com/in/sankalpthakur/ Hugging Face: https://huggingface.co/sankalpsthakur Scripts generated with Claude Opus. Synthetic narration generated with Edge TTS. Reviewed, listened to, and approved by Sankalp Thakur.

    Context Is the Bottleneck (Harness Design, Part 2)

  4. Oct 12

    The Loop and the Hands: What a Coding Harness Is (Harness Design, Part 1)

    Platform engineer Ines Varga doubles her coding agent's success rate without changing the model, and her mentor Gideon Achterberg uses it to explain the harness: the loop, the tools and the guard rails around a model with no memory. We trace the history from ReAct to SWE-agent's agent-computer interface, tour OpenHands, Aider, mini-swe-agent and the commercial agents, then go inside the harness from 'An Empirical Study of Harness Design for Coding Agents' (arXiv:2609.20804). In that study, predefined tools scaffold weaker models (worth 15 points on SWE-Bench Verified for Nemotron-3 30B), while a bash-only interface made Nemotron-3 550B slightly more accurate at half the cost, and Mistral Medium 3.5 needed tools for repo editing but preferred the shell on Terminal-Bench. The episode also covers granularity, tool design, how to read the statistics, safety gates, stuck detection and sandboxes. Chapters 00:00:00 Cold open: same brain, different hands 00:03:57 First principles: a model is a function, a harness makes it an agent 00:08:04 A short history: from reasoning traces to agent-computer interfaces 00:12:12 The agent-computer interface, and why models need their own 00:15:40 The harness family tree, in twenty twenty-six 00:19:50 Inside the paper's harness, piece by piece 00:23:24 The tools on the bench 00:26:25 The action-space result: scaffolding for some, overhead for others 00:32:44 Granularity: how the hands shape the code 00:36:06 One ticket, two pairs of hands 00:39:47 Designing a tool a model can actually use 00:43:06 Reading the action-space numbers with care 00:47:03 Gates and guardrails: what the paper held still 00:50:17 Sandboxes: where the hands are allowed to reach 00:52:22 What this means, and where we go next Created by Sankalp Thakur. A Transient Labs production. Website: https://transientlabs.ai GitHub: https://github.com/sankalpsthakur X / Twitter: https://x.com/sankalpsthakur LinkedIn: https://www.linkedin.com/in/sankalpthakur/ Hugging Face: https://huggingface.co/sankalpsthakur Scripts generated with Claude Opus. Synthetic narration generated with Edge TTS. Reviewed, listened to, and approved by Sankalp Thakur.

    The Loop and the Hands: What a Coding Harness Is (Harness Design, Part 1)

  5. Oct 12

    The Brain and the Frontier: Posner Molecules, Microtubules, Diamond Thermometers and What Would Settle It

    Day three, in a storm, with a bottle of whisky riding on the outcome. Gus teaches the three ways to estimate decoherence times (scattering, spectral density, spin relaxation) and a fourth way: test the consequence, not the coherence. Wren presents Fisher's 2015 nuclear-spin proposal, the 2018 Player and Hore critique, the 2025 PNAS lithium-isotope calcium phosphate result and a 2025 negative model. They weigh Orch-OR against Tegmark (2000), the Gran Sasso exclusion of parameter-free Diósi-Penrose collapse (Donadi et al. 2020; Derakhshani et al. 2022), Franks and Lieb on anaesthetic targets, Kerskens and López Pérez (2022) and its 2023 rebuttal, Babcock et al. (2024) superradiance, Wiest et al. (2024) epothilone B, the 2018 xenon isotope study, and Zadeh-Haghighi, Craddock, Simon and colleagues' magnesium-25 tubulin result (Science Advances 2026). Then NV-diamond thermometry (Kucsko 2013), Chrétien's 50 °C mitochondria and Baffou's heat-conduction objection, the episode 89 preprint graded, a funding list of decisive experiments, and the final ledger. Wren Halloran and Augustin "Gus" Ferreira are fictional. Chapters 00:00:00 Cold open: the storm and the wager 00:01:51 How decoherence times are estimated 00:06:58 Why the brain is the hardest place to look 00:10:23 Fisher's Posner molecules: a quantum brain built from nuclear spins 00:14:38 Testing the Posner idea 00:18:33 Orch-OR: consciousness, collapse and microtubules 00:21:39 The critics: Tegmark, Gran Sasso and the anaesthesia question 00:25:57 An isotope puzzle in anaesthesia 00:28:17 New evidence, twenty twenty-two to twenty twenty-six 00:35:58 Quantum sensing inside cells: diamonds as thermometers 00:41:19 The mitochondria claim, placed on the ledger 00:44:25 What experiments would settle things 00:47:25 The final ledger and the wager Created by Sankalp Thakur. A Transient Labs production. Website: https://transientlabs.ai GitHub: https://github.com/sankalpsthakur X / Twitter: https://x.com/sankalpsthakur LinkedIn: https://www.linkedin.com/in/sankalpthakur/ Hugging Face: https://huggingface.co/sankalpsthakur Scripts generated with Claude Opus. Synthetic narration generated with Edge TTS. Reviewed, listened to, and approved by Sankalp Thakur.

    The Brain and the Frontier: Posner Molecules, Microtubules, Diamond Thermometers and What Would Settle It

  6. Oct 12

    Spins and Tunnelling: The Robin's Compass, Cheating Enzymes and the Protons in Your DNA

    Day two on the houseboat. Wren and Gus build the radical-pair mechanism from first principles, put a quantum compass through an engineering design review, and follow robin Cry4 from candidate (Günther 2018) to magnetically sensitive protein (Xu et al., Nature 2021), plus the 2023 to 2025 follow-ups (tryptophan-chain timing, Cry4b ruled out, membrane ordering, superoxide and quantum Zeno models). They weigh the radiofrequency disruption results (Ritz 2004, Engels 2014, Goforth et al. 2025 sea turtles), the magnetite beak story's collapse, the 100,000-fly null result (Bassetto 2023), monarch cryptochrome knockouts and the neural recordings due after December 2026. The second half covers tunnelling from first principles, soybean lipoxygenase's kinetic isotope effect of about 80, the Klinman versus Warshel dynamics debate, Turin's vibration theory of smell, and Löwdin's DNA tautomers, from Slocombe, Sacchi and Al-Khalili (2022) to the 2025 counter-analysis. Wren Halloran and Augustin "Gus" Ferreira are fictional. Chapters 00:00:00 Cold open: a robin in a funnel 00:02:26 The weakness of the Earth's field 00:04:06 Radical pairs from first principles 00:08:36 A design review for a quantum compass 00:12:44 Cryptochrome: from candidate to cryptochrome four 00:18:04 The radiofrequency fingerprint 00:23:01 Fights, failures and a fruit fly 00:26:43 Magnetic field effects beyond the compass 00:29:36 Tunnelling from first principles 00:33:42 Enzymes that cheat at chemistry 00:39:03 A cautionary tale: the quantum nose 00:42:57 The protons in your DNA 00:47:23 The ledger after day two Created by Sankalp Thakur. A Transient Labs production. Website: https://transientlabs.ai GitHub: https://github.com/sankalpsthakur X / Twitter: https://x.com/sankalpsthakur LinkedIn: https://www.linkedin.com/in/sankalpthakur/ Hugging Face: https://huggingface.co/sankalpsthakur Scripts generated with Claude Opus. Synthetic narration generated with Edge TTS. Reviewed, listened to, and approved by Sankalp Thakur.

    Spins and Tunnelling: The Robin's Compass, Cheating Enzymes and the Protons in Your DNA

  7. Oct 12

    Light and Coherence: What Photosynthesis and Vision Really Tell Us About Quantum Biology

    A sceptical retired NMR physicist and a young ultrafast spectroscopist start a graded "ledger" on a houseboat lab, sorting every quantum-biology claim into established, strong evidence, contested, speculative or fringe. Day one covers light: how antennas funnel excitons, the Engel et al. 2007 Nature beats at 77 K, environment-assisted quantum transport, Duan et al. 2017 (electronic coherence gone in ~60 fs), the 18-author 2020 Science Advances reassessment, Brumer's sunlight objection, the 2023 single-photon LH2 experiment, the 2026 reviews, spectral densities and decoherence estimates, Marcus electron tunnelling, and the vibrationally coherent isomerisation of retinal (Schoenlein 1991, Wang 1994, Johnson 2015, Gruhl 2023), plus single-photon vision and a checklist for reading quantum-biology headlines. Wren Halloran and Augustin "Gus" Ferreira are fictional. Cross-references episode 89 (the Fudan mitochondria preprint). Chapters 00:00:00 Cold open: the houseboat and the ledger 00:03:09 What it would even mean for biology to be quantum 00:07:32 Catching sunlight: antennas, excitons and the FMO complex 00:11:33 Rings of light: delocalisation you can see 00:14:49 Two thousand and seven: the beats that launched a field 00:17:50 Noise as a helper: environment-assisted transport 00:20:54 The reassessment: vibrations, sixty femtoseconds, and eighteen authors 00:26:18 What survives: vibronic mixing, single photons and the twenty twenty-six view 00:30:43 Decoherence by the numbers: the spectral density 00:33:46 Electron tunnelling: the quiet quantum effect everyone accepts 00:36:49 Vision: the fastest twist in biology 00:42:45 Can you see a single photon? 00:45:39 Lessons for engineers: what the reassessment teaches 00:48:10 How to read a quantum biology headline 00:50:09 The ledger after day one Created by Sankalp Thakur. A Transient Labs production. Website: https://transientlabs.ai GitHub: https://github.com/sankalpsthakur X / Twitter: https://x.com/sankalpsthakur LinkedIn: https://www.linkedin.com/in/sankalpthakur/ Hugging Face: https://huggingface.co/sankalpsthakur Scripts generated with Claude Opus. Synthetic narration generated with Edge TTS. Reviewed, listened to, and approved by Sankalp Thakur.

    Light and Coherence: What Photosynthesis and Vision Really Tell Us About Quantum Biology

  8. Oct 12

    Light in the Powerhouse? The Quantum Mitochondria Preprint

    Reliability engineer Theo Marchetti gets a screenshot from his sister Ines, a mitochondria biologist in Lisbon: "scientists proved every mitochondrion is a tiny quantum device running at 71 THz." Over a week they take the paper apart together. We build the background from first principles: mitochondria and the electron transport chain, Peter Mitchell's chemiosmosis and the ATP synthase turbine, and whether life's energy efficiency really needs quantum physics. Then we look at the history of quantum biology: the 2007 photosynthesis coherence result and its 2020 reassessment, the robin's radical-pair compass, enzyme tunnelling, Fisher's Posner molecules, Penrose–Hameroff and the "warm, wet and noisy" decoherence objection. Next comes a pendulum-and-tuned-mass-damper primer on polaritons, followed by the paper itself: its FTIR spectra, its cavity model and its laser–ATP experiment. Last, a reliability engineer's stress test, an account of how the paper was received (careful news coverage, wild hype on X, no expert review or replication yet), every claim sorted into one of three tiers, and the specific experiments that would confirm or refute it, including a carbon-13 isotope test. Chapters 00:00:00 Cold open: the screenshot at the turnaround 00:03:14 The powerhouse: what a mitochondrion actually is 00:05:54 Chemiosmosis: the hydro dam inside the cell 00:09:45 The efficiency puzzle, or is it? 00:13:56 Quantum biology: big claims and their afterlife 00:22:03 Polaritons: when light and matter hold hands 00:26:29 The anomalous peak 00:30:41 The model: a mitochondrion as a resonant cavity 00:36:02 The light experiment: ten percent more ATP 00:38:18 Stress test one: is the ghost peak carbon dioxide? 00:41:26 Stress test two: what changed between living and dead? 00:44:09 Stress test three: does the cavity hold water? 00:49:14 Stress test four: the ATP experiment 00:53:08 How the paper was received 00:56:32 The ledger: three bins 00:58:39 What would settle it 01:00:43 Closing: the peak that wasn't a bearing Created by Sankalp Thakur. A Transient Labs production. Website: https://transientlabs.ai GitHub: https://github.com/sankalpsthakur X / Twitter: https://x.com/sankalpsthakur LinkedIn: https://www.linkedin.com/in/sankalpthakur/ Hugging Face: https://huggingface.co/sankalpsthakur Scripts generated with Claude Opus. Synthetic narration generated with Edge TTS. Reviewed, listened to, and approved by Sankalp Thakur.

    Light in the Powerhouse? The Quantum Mitochondria Preprint
See All (96)

About

Long-form engineering, computing and mathematics. Created by Sankalp Thakur. A Transient Labs production. Website: https://transientlabs.ai/the-deep-end GitHub: https://github.com/sankalpsthakur X: https://x.com/sankalpsthakur LinkedIn: https://www.linkedin.com/in/sankalpthakur/ Hugging Face: https://huggingface.co/sankalpsthakur Scripts: Claude Opus. Synthetic narration: Edge TTS. Reviewed, listened to and approved by Sankalp Thakur.

Information