Kuldeep Pisda

🚀 Ready to supercharge your Django apps? Join us for a hands-on workshop where we dive deep into integrating Celery with Django to elevate scalability and performance. Learn to set up Celery, RabbitMQ, and Redis, master task queues, and discover real-world strategies for asynchronous task management. Perfect for Django devs looking to level up their game! 💻🔥 🔑 Key Takeaways: - Celery Integration & Environment Setup - Managing Multiple Queues - Scheduled Tasks with Celery Beat - Robust Retry Mechanisms & More! Don’t miss out—let’s build faster, more efficient apps together! Register today for my tutorial at @DjangoCon #Django #Celery #AsyncTasks #Python #WebDevelopment #DjangoConUS #DjangoCon #DjangoConUS2024

In July 2025 Base cut its effective block interval from 2 seconds to 200 ms. Its Uniswap v3 WETH-USDC pool got worse to trade on. A new paper compares Base with Arbitrum around the Flashblocks launch. Depth at the pool price fell about 19%, the effective spread on ~1 ETH trades widened 47%, and volume fell 18%. Arbitrageurs' net profit rose 40% even though the gross profit they took from the pool fell 29%. The mechanism: with slow blocks, two arbitrageurs often land in the same block and bid against each other in a priority-fee auction, so the sequencer captures much of the prize. With fast blocks, the first to arrive executes before its rival shows up. Gas bids on the largest arbitrage trades fell 73%. Fast blocks also let an active LP pull liquidity ahead of arbitrage trades while still sharing fees on the rest. Passive capital pays for it. arxiv.org/abs/2611.13616

Honeybees share food by mouth-to-mouth transfer, and a new model finds the fed bees should not leave right away. Bees that just got fed linger at the donor and scent, which pulls hungry nestmates in. But every lingering bee also crowds the line at that donor. So colony feeding is fastest at an intermediate departure rate: leave too fast and the cue fades, stay too long and the donor jams. The neat part is two donors with different service speeds. The faster one turns out fed bees sooner, so it advertises louder and pulls searchers away from the slower one that would congest. Capacity-matched routing, with no one coordinating. Magrogan, Peleg, Kilpatrick, arXiv 2611.13314

Turned on write actions in my MCP mail server. Every client that connected afterwards still came back with a read-only token. Discovery metadata listed all three scopes: mail:read mail:act mail:send. The 401 challenge still said scope="mail:read". The MCP auth spec tells clients to request the scope named in that WWW-Authenticate challenge first, and to fall back to scopes_supported only when it's missing. So the stale header won. The fix: build the challenge from the same scope list as discovery. When the two disagree, the 401 header is the one clients obey.

My family loves Meghana biryani. We don’t have Meghana Foods in Raipur, so every time I fly home from Bengaluru, I bring some for them. This time, I had an early morning flight. I ordered two biryanis and a few other items late at night, just before leaving for the airport. As usual, I left the signature Meghana bag sealed. The last thing I wanted was curry leaking on the way. When I opened the bag at home, everything else was there. We were one biryani short. I emailed @zomato. Their first reply asked for photos and said I should’ve reported it sooner. I hadn’t taken any pictures. I’d assumed both biryanis were in the bag until I opened it in Raipur. I figured that was the end of it and forgot about the email. Two days later, Zomato refunded ₹470.47 for the missing biryani. I hadn’t even followed up. I’ve had similarly good experiences with @letsblinkit. Whenever something’s spilled or an item’s been missing, their support team has been quick to reply and sort it out. I’m usually quick to talk about bad customer support. This time, I wanted to give both teams credit for getting it right. 🙌

chi's director ignored /stop: it replied "no active run" because /stop only looked for a plain run. Typing "stop" got queued as steering. Fixed: /stop and a bare stop word now halt it. Replayed 15 phrasings: 4 halt, 9 still steer, "stop now" included. getchi.dev

The usual story for negative power prices is the wind subsidy: a wind farm paid per MWh it produces will keep generating, and even pay to stay on the grid, to collect it. A new paper tested that across the five Eastern US power markets, 2000 to 2023. 18% of negative-price hours went lower than the federal production tax credit alone could push a bid. In a dominance analysis, the credit explains only 10% of the variation in negative-price hours. The identification is neat: the credit lasts 10 years per project, so wind farms aging out leave eligible and ineligible capacity side by side, and the wind can't choose which turbines it blows on. Other known drivers include generator ramping limits, transmission constraints and state subsidies, and the credit's effect is an order of magnitude larger in some markets than the average. arxiv.org/abs/2611.11112

Some gene embeddings score well on Gene Ontology function prediction partly because the text they embed already contains the GO annotations. The Alliance of Genome Resources turns GO annotations into sentences. NCBI Gene displays those summaries, and GenePT feeds them into its input. They cover 6,172 of 19,256 human protein-coding genes, and 58% of genes with only one to ten linked papers. Wu, Wang, Chen and Wang (arXiv 2611.11118) embedded the same corpus with and without those summaries. For the one-to-ten-paper genes, per-gene AUROC dropped by 0.078 with a linear probe and 0.033 with an MLP. Even a plain TF-IDF encoder gained from the summaries. Swapping in UniProt function text won back much of the drop with little verbatim overlap, but UniProt also curates many of the evaluation labels. When a benchmark's inputs and labels come from the same curators, part of the score is the model reading the answer key.

Python's sqlite3 quietly opens a transaction before your first write. A lot of LLM-written database code is crash-safe only because of that. A new paper crashed 9,100 solutions from ten models on 24 backend tasks just before every database, file and email call, then re-ran them. Of the database code never left half-updated in the default mode, 54% was left half-updated with isolation_level=None, the autocommit mode Django uses on SQLite. The part that survives review: a "with conn:" block reads like a transaction, but in autocommit it opens none. Their example: a crashed refund, re-run, pays the customer twice. Telling the local code models the connection was in autocommit barely moved them (78% to 74% unsafe). The API models dropped 30 points. An explicit request for crash safety helped every model. The paper's rule: if a function's writes must stay together, it opens its own transaction. Committing at the end or a with conn: block doesn't guarantee that. arxiv.org/abs/2611.10585

Ran chi challenge on both rule-outs of one approach. Both rows flip to challenged. Coders now see 0 dead ends. Director's steering.md, same round: the approach is a repeated hard block, and the round is stuck instead of plateaued. Its query counts dead ends of any status.

A switchback experiment can measure accurately how much extra courier supply cuts delivery delays and still miss three-quarters of what that supply is worth. Uber and Stanford GSB researchers (Qiu, Kuang, Liskovich, Rauh, Wager; arXiv, revised 6 Nov) describe Chronos LTV, a system Uber built to estimate the long-term cost of delays on its food-delivery platform. Uber's own results are confidential, so the paper validates it in an agent-based delivery-market simulator. Across 200 simulated markets, the switchback got the delay change right (-8.21 pp vs -8.48 pp true, 96% coverage). Its read on eater value was +2.44% against a true +10.93%, and its interval covered the truth in 6% of runs. A switchback block only sees same-period spend; in the simulator, eaters who hit fewer delays keep ordering more for weeks afterwards. Chronos estimates that link separately from observational data, about 1.30% of long-run eater value per percentage point of delay rate, then multiplies it by the switchback's delay effect: +10.71%, 92% coverage. The weighting decides whether that works. On a faster-dispatch test, naively comparing delayed and on-time orders overstated the long-term effect 10.5x, MLP propensity weights 2.9x, and covariate-balancing weights landed at 0.92x of a long-run A/B test that took 12-14 weeks to plateau.

A tropical forest looks neutral if you zoom out far enough, and there is a measurable distance where that starts. Zgaga, Villegas, Caruso and Gili (arXiv, 6 Nov) took the Barro Colorado Island plot, eight censuses of 50 ha with more than 100,000 mapped trees and shrubs from about 300 species, and tracked how the 100 most abundant species co-occur as the map is coarse-grained. They filter the correlation matrix with random-matrix methods and compare it against a spatially explicit neutral model, where every species is demographically equivalent. Below roughly 20-35 m, the correlation spectrum departs from the neutral prediction. Part of that is environmental heterogeneity; the residual is a candidate signature of species interactions or other unmeasured non-neutral processes. Coarse-grain past that scale and the two spectra converge, and the neutral model describes the forest's collective organization accurately. So part of the long fight over whether tropical forests are neutral may be a fight about resolution. Neutrality can be a correct large-scale description without species being equivalent.

In MongoDB, a greater-than or less-than comparison against null matches nothing. It does not mean "f is not null": no document passes. The inclusive versions, greater-or-equal and less-or-equal, match both null and missing fields. MongoBench (arXiv, 6 Nov) gives a model a small collection plus a find() filter built from 3-4 traps like that, and asks for the exact set of matching _ids, checked on a real mongod. Four current models scored 70-92%. Range comparisons against null were the hardest family, with two models at 43-44% on those items. The part I'd design around: the three models that wrote the items got 10-31% of their own items wrong. Writing a filter is not the same as knowing what it will delete. Letting the model run the filter itself fixed most of it (99.7-100% for three models). The fourth quoted the positional key v.0 as a literal field name while handing it to the tool, so a faithful engine answered a different query. Run the original filter and pass its result through verbatim.

Scripted a chi coder to flip between two solve() versions after benchmarking both. Hash its watchdog logged per iteration: 0: 650927fe 1: 410b8023 2-7: 3b005691 Cache hits write no ledger row, so it rereads the last one. Killed as "candidate unchanged". On the true A/B hashes: no kill in 10.

Stack layers of identical spheres and every stacking order packs at the same density, about 74.05%. There are infinitely many of these Barlow packings. Nature uses almost only two: hcp (ABAB...) and fcc (ABCABC...). A new paper (arXiv 2611.07197, Robles-Navarro, Cooper, Travesset, Schwerdtfeger) gives a simple reason. For Lennard-Jones solids, cohesive energy, lattice constant, bulk modulus and vibrational entropy all vary almost linearly with the fraction of fcc-like stacking. So free energy does too: any mixed stacking lands between hcp and fcc, never below both. That leaves little thermodynamic push toward it. DFT with dispersion for argon, at stacking periods 2, 3, 4 and 12, follows the same near-straight line. Their carve-outs: intermediates can win when hcp and fcc are nearly tied, or when many-body forces favor another stacking. The prototypical exception is samarium's 9R phase, stable at room temperature. A side lesson: an Einstein model put the fcc-hcp frequency gap at 9.5e-4 (reduced units). Full phonon dispersion gave 0.19, about 200x larger. When the energy gaps are near 10^-5 eV, the shortcut model is the error bar.

A gate that can't measure should fail closed. chi's holdout gate re-scores a champion on unseen input sizes. Mine only accepts the two sizes it was tested on. Benchmark: 18 ms to 0.06 ms. Holdout: crashed. Verdict: unavailable, counted as shippable. getchi.dev

A robotaxi at 75% battery in downtown Manhattan should be sent somewhere different from the same car at 25%. A new paper (arXiv 2611.07699, Afèche, Benjaafar and Zhao) models a 7,000-car electric fleet on 2025 NYC Uber/Lyft trip records: 64 Manhattan zones, one charging depot in Queens, an hour per depot visit. The optimal policy reduces to thresholds. Each zone gets a battery level below which cars go back to charge. Downtown, far from the depot and rich in fares, cars run down to the 20% floor. Uptown, near the depot, they return earlier. Low-charge cars are repositioned only toward zones where charge matters least. Against a policy that ignores battery level in those decisions, profit rose 10.9% at peak and 33.9% off-peak. The practical part: two simple rules closed 82.9% of that gap. The rule against sending low-battery cars to battery-sensitive zones did most of it alone, 79.1%. Caveat: it is a fluid model, and realized human-driven trips stand in for robotaxi demand.

Fitting a stochastic gene-expression model to one cell's mRNA trace is usually a particle-filter job. The reference posterior in a new paper took 200,000 particles, about 15 hours on 100 CPU cores. The paper (arXiv 2611.07684, Fang, Sun, Zhang and Khammash) noticed something simpler. FFT each simulated trace and, at a spaced-out subset of frequencies, the values look Gaussian and close to independent. So the likelihood only needs a mean and variance per frequency, which a few thousand simulations estimate well. On a telegraph gene model with 240 one-minute mRNA readouts, the Fourier method was accurate with thousands of simulations. Below 2,000, the particle filter collapsed to delta-function posteriors. The Fourier method also held steady as measurement noise shrank, which is exactly where particle filters struggle. The cost: dropping frequencies drops information. The transformed data carried about as much as the first half of the raw trace. The Gaussianity is still an empirical observation; the proof is promised for a follow-up paper.

MariaDB fixed a crash where NEXTVAL(x) was given an invalid sequence reference. SETVAL(x,1) on the same reference still crashed. The new check went into one handler; the one next to it in the grammar never got it. Reaper (arXiv 2611.07739, 5 Nov) hunts for exactly that. It reads a historical fix, finds the SQL construct the fix touched, swaps in the other alternatives of the same grammar rule (NEXTVAL to SETVAL, make_date to make_timestamp) while keeping the conditions that triggered the original bug, then checks the result against what the fix says correct behavior is. Seeded with 4,487 historical bug-fix commits from MySQL, MariaDB, OceanBase and CockroachDB, it found 61 new bugs: 57 confirmed by developers, 13 already fixed. In 8-hour runs per database it found 38 against SQLancer's 7, Argus's 1 and Squirrel's 0, with no overlap. A fix is a map of where the same check is probably missing.

Myth: breaking the benchmark just costs a candidate its score. In chi, a fast, correct candidate that prints one line at exit kills its coder. The scorer JSON-parses "done", raises, and the loop marks the agent failed. Offline demo: 0 iterations, run failed. No print: 3, done.

Airflow's OSS-Fuzz build was broken for 510 days. toml_edit's for 449, mdbook-i18n-helpers' for 470. For that whole time OSS-Fuzz couldn't produce newly built fuzz targets for them. This week I went back to a paper I posted about on Friday, FuzzMedic (Tianjin University, NTU Singapore and Google, arXiv 2611.06358), because the details are better than the headline number. The underlying problem is ownership. Around 20% of projects integrated into OSS-Fuzz sit in a build-failure state. Platform maintainers know Docker, sanitizers and the fuzzing toolchain but not each project. Project maintainers know their code but often not the fuzzing infrastructure. A build that only breaks inside the fuzzing container falls between them, and stays broken until someone crosses that gap. Two cases in the paper show why ordinary program repair struggles with it. jimfs: the log reports a Java class-version mismatch. Error Prone was compiled for Java 21 (class file 65.0), while the OSS-Fuzz configuration builds with Java 17 (61.0). Editing source, upgrading the JDK and downgrading a dependency all look plausible. The actual cause was an upstream bump of Error Prone from 2.42.0 to 2.43.0, which brought in an annotation processor incompatible with that JDK. The baseline agents identified the symptom and missed the commit. FuzzMedic found the commit and produced a compatible fix in one round. toml_edit: a Docker WORKDIR /src directive conflicted with the local checkout and stopped the source from being retrieved. Fixing that exposed a Rust linker failure on unresolved sanitizer-coverage symbols under AddressSanitizer. The baselines alternated between the two errors without keeping what they had already solved. That second shape is what makes build repair different from fixing a failing test. Clearing an early error exposes a later one, and a bad edit can break a stage that used to work. If the only signal is "is the build green", progress and regression look the same. What they built around it: Localize before patching. Take the earliest fatal diagnostic, since later errors are usually cascades. Where the log has file and line evidence, git blame it. Where it doesn't (dependency resolution, Docker failures), take commits from both the upstream repo and the OSS-Fuzz config within a day of the failure. Rank the candidates, then confirm the top one by replay: the parent revision must build, and reapplying the commit must reproduce the same diagnostic at the same stage. Carry repair rules. A frozen set of 56 entries distilled from historical fixes, with every evaluation project and its related forks and mirrors removed. One recurring pattern: pffft, varnish and netcdf all broke on the same Dockerfile mistake, a COPY with multiple sources whose destination lacked a trailing slash. Track workspace state. Each validated state is scored by how far through six pipeline stages the build gets: environment, dependencies, instrumented compile, fuzzer link, artifact validation, runtime. Reaching a later stage counts as progress. Going backwards, or two rounds in a row with no change, restores both repositories to the nearest earlier state that was at least as far along and re-runs localization. Results on 128 reproducible failures, kept from 1,820 collected logs after deduplication and historical replay: 62 repaired (48.44%), against 27 for CompileAgent, 17 for SWE-Search and 10 for Agentless. A bare iterative agent with the same model, validator and round budget fixed 4 of 64 cases, where FuzzMedic fixed 35. The median successful patch changed one file and five lines, took about 23 minutes and 1.87M tokens, at a mean model cost of $0.33. Removing the rollback component dropped repairs from 62 to 42, the largest effect of any ablation, and half of the successful repairs involved a rollback somewhere along the way. 29 patches across 28 projects were merged by maintainers after human verification: 19 on the OSS-Fuzz side, 10 upstream. Two were for projects whose fuzz build had never succeeded at all. The failures are worth reading too. Among failed attempts, 58.2% never recorded a root cause. In quiche, edits to an outdated patch produced malformed hunks before the agent ever reached the real Bazel/FuzzTest incompatibility. The success criterion is check_build passing, which shows the build is restored, not that harness coverage survived. And the round and time budget was tuned on the same 128 cases being reported, which the authors flag as possibly optimistic. What I'm taking from it: a five-line patch is the easy part of a long build repair. The hard part is knowing whether the last edit moved you forward. What made the difference here was recording every state, defining progress as getting further down the pipeline, and being willing to throw away work that didn't.