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

Myth: reacting faster to a stuck agent reaps it sooner. chi now nudges a coder on its first timeout with no result. Replay, same candidate timing out every round: old watchdog killed it on round 6, new on round 10. The nudge outranks the repeat kill. 600s a round: +40 min.

Of 66 published techniques for making LLM-written code cleaner, faster or more secure, one beat simply sampling about four answers and keeping the best. Prashank Kadam (arXiv, 28 Oct) re-ran all 66 on one model, Gemma 4 31B, against that baseline: best-of-n, ranked by Pylint, Bandit and Safety findings. 51 of the 66 improved their targets over a single call. Against best-of-n, 1 came out ahead beyond the 95% interval (CodeChain) and 22 came out below it. With 66 comparisons, about two such wins are expected by chance. In the paper's matched-budget grid, when prompting, planning and retrieval ran at four or more calls, their gains came from the shared ranker, not the instruction, the plan or the retrieved text. So a reported gain needs a budget-matched sampling baseline, scored on a measure the technique wasn't optimizing.

Train a small chemical language model on 1,000 molecules and it turns into a lookup table: it regenerates 99.98% of its training set exactly, and only 15 of every 10,000 samples are structurally new. A Michigan Tech paper (arXiv 2610.38835) fixed that by writing each training molecule many ways. In SAFE notation a molecule is a string of fragment blocks, so you can shuffle the block order and also respell the atoms inside each block. With both kinds of shuffling, structurally new molecules rose to 7,079 per 10,000. The surprise is that the two axes don't add up. At the same 32,152 training strings, atom order alone gave 712 new molecules per 10,000, fragment order alone 1,026, both together 6,977. With one axis varied, the model still regenerated about 90% of its training set exactly. Randomized SMILES, the usual trick, got 2,562 at matched size. A third of its novel molecules were near-copies of a training molecule, against 5.5% for SAFE.

On Manifold, a market priced at 50% on someone's own goal ("Will I...") resolved YES 39% of the time. At the same price, markets on outside events resolved YES 47% of the time. Chen, Zheng and Zhu (arXiv, 30 Oct) reconstructed every order in 6,150 markets where 1,660 users asked about their own futures. Those resolved YES 4.4 points less often than their final prices implied. Markets on outside events were off by 0.9. The optimism came from trading. Almost every market opens at the default 50%. On day one, other traders pushed outside-event markets 4.5 points toward NO, the usual base rate. Most self-markets moved only 0.8. The correcting information arrived late and from the person. Their own orders supplied 38% of the price movement toward the eventual outcome, against 9% from creators of outside-event markets, and nearly half of it came in the last day. That's why fitting a correction barely helps: trained on earlier markets, it cut later Brier scores by only 1.5%. The prices weren't just shifted toward YES. They were worse at telling who would succeed.

chi's director can stop itself at a target score. The check reads the raw best benchmark, not the noise guard's verdict. Replay, target 500, baseline 636: one lucky 480 sample. Guard re-runs it 3x, median 640, round demoted to plateaued. Director halts anyway: target reached.

In a lab invite system I'm building, the Copy link button on an invite turned out to be a way for an admin to make themselves an owner. The rules looked tight. A lab has four roles: owner, admin, member, viewer. Owners and admins manage people. When you send invites, the server rejects any row whose role ranks above yours, so an admin who tries gets "You can't invite someone as owner." Promoting someone to owner, demoting an owner, or removing one also needs an owner. The gap was in what happens after an invite exists. On the team page, owners and admins could do three things to any open invite: copy its link, resend it, or revoke it. Those endpoints checked one thing, whether you manage this lab. They never looked at the role on the invite. That mattered because of how the link works. Open it signed out, with no account on the invited address yet, and it creates the account from a name and a password, then puts that account in the lab with the invite's role. Holding the link is the proof. Nothing else checks that you own that email. So an owner sends an owner invite to a new collaborator. An admin clicks Copy link on it, signs out, opens the link, picks a password, and now controls an owner account in the lab under the collaborator's address. Copying also issues a fresh token, so the link the collaborator got by email stops working. Two paths were already closed. Signed in as themselves, the admin would have been stopped, because accepting while signed in requires the account's email to match the invite. And if the invited address already had an account, signup returned a 409 and asked them to sign in. The hole was the new-account path, which is the usual case for an invite. The fix moves the rank check from sending to handling. Before copy, resend or revoke, the server compares the invite's role rank with the actor's and returns 403 if the invite outranks them: "Only an owner can manage an owner invite." A parametrized test has an admin try each of the three actions on an owner invite, expects 403, and checks the invite is still open. A second commit hides those buttons from admins on owner invites, but the server check is the one that counts. Both commits landed about five minutes after the feature commit, on the feature branch, before it reached main. What I took from it: a check on who can create something says nothing once it exists. If an object carries authority, like a link that grants a role, then any action that reveals or reissues it is the same as granting that authority, and it needs the same check.

A single superconducting vortex hopping between two pinning sites dissipated less energy per cycle than Landauer's limit, kT ln 2. That isn't a violation, and the reason is the interesting part. Fridman et al. (Hebrew University, arXiv, 27 Oct) drove one vortex in a 55 nm niobium film at 4.2 K with an oscillating current and read the work per cycle from the out-of-phase part of its motion, imaged with a SQUID-on-tip. When the drive is slower than the vortex's own thermal hopping rate (3-7 kHz in the wells where this showed up), the k ln 2 of entropy handed to the bath while tilting it into one well comes back as it randomizes again. But a bit that randomizes that fast can't store anything. In their simulation, erasing a bit that actually holds its state costs about 5x the limit at 150 kHz.

A runtime invariant detector that lets an LLM write the invariants had about a quarter of the false-positive rate of the template-based classic. Oca (Arvan, Ernst, Kellogg, arXiv, 30 Oct) asks gpt-4.1-mini for candidate pre- and postconditions on Java methods and keeps the ones no test run falsifies. In a blinded sample of 50 surviving invariants per tool, 18% of Oca's were false positives vs 76% of Daikon's. The awkward part: every candidate is code running inside your program, so it can mutate state or hang. Oca doesn't try to prove purity. When a test newly fails, delta debugging finds the guilty invariant and disables it, so purity holds only as far as the tests look. On 25 Defects4J bugs, Oca's invariants exposed 9, Daikon's 4.

Software proposes gut metabolite structures faster than any lab can confirm them, and confirming one means having the actual molecule. Koyama (arXiv, 29 Oct) sorted 10,645 inferred structures. 7,241 can be bought, but what's for sale is what has already been studied.

chi's director once spun 26+ rounds in 6s, asking its strategy brain for a plan each round with nothing new to plan on. Replay with an instant stub brain, empty slices: old loop planned 76,513 times in 6s. Now once. It replans only when something moved. getchi.dev

In DoorDash's A/B tests, the treatment effect is a steadier signal than the metric it moves. A new paper from Stanford and DoorDash authors looked at 196 DoorDash experiments. The daily effect was far less volatile than the control-group metric, and its trend differed from the metric's (p = 2.3e-5). Control-group levels added nothing once the effect series was known (p = 0.36). So their method predicts the day-55 effect straight from the day 6-19 effect series, fitted on the most similar past experiments. Against an armwise surrogate baseline, RMSE fell by about 15% to 42% depending on archive size, using only aggregate daily numbers. That makes an archive of finished long-running tests worth keeping, even when each one only answered its own question.

In a simulated quantum circuit, an error-detection check broken by a T gate does more harm than no check at all. Ignoring its result afterwards makes it worse. A new arXiv paper (Chongder & Kumar) turns this into a Qiskit game: 6 data qubits, depth 18, three spacetime checks, 16 T gates, scored by density-matrix simulation. No checks: fidelity 0.628. T gates placed where they commute with the checks: 0.892 after post-selection. T gates placed blind: the checks fire on perfect runs, 17.4% of shots survive, fidelity 0.158. The obvious fix is to stop listening to the broken checks. That gives 0.102. The ancilla is still entangled with the data, so throwing away its outcome is a measurement you forgot: it dephases exactly the coherence the T gates were added to create. A broken check has to come out of the circuit, not just out of the post-processing.

In Maven, the patched release usually ships before the CVE goes public. For about 82% of 832 CVEs tracked in a new study, the fixed version was out before NVD disclosure. Downstream projects still waited. 85.3% of 1,677 observed adoptions came after disclosure, and in a sample of the linked PRs and issues only 7% mentioned the vulnerability at all. Most fixes arrived as ordinary dependency bumps. So routine updating is doing the security work. Dependabot version updates were associated with a median 10.4 days from fixed release to adoption, vs 110.6 days without. Its security updates only open PRs for known vulnerabilities. Yet version updates were configured for just 8.5% of adoptions, though 78.2% of the repos used Dependabot at some point. (arXiv, 27 Oct, Hasan, Pan, Tian, Guizani, Ding; Queen's and McGill)

hyperexponential raised $91M from investors. Guidewire just agreed to pay $1.7B in cash for it, about 19x every disclosed dollar that went in. The deal, announced 27 October: - $1.7B base price, plus up to $400M tied to recurring revenue targets - about 23x the money raised if that earnout pays in full - $1.0B funded with new term loans, the rest from cash on hand hyperexponential is a London company selling AI underwriting and pricing software to commercial insurers. Per Tracxn, it raised an $18M Series A in 2021 and a $73M Series B in 2023. It was already a Guidewire alliance partner. So a core-system vendor for insurers is borrowing more than half the base price to own a tool its customers had already plugged in. The deal hasn't closed yet; Guidewire expects that in Q2 of its fiscal 2027. For vertical software startups, the strongest acquisition pitch may be the integration you already shipped.

Wrote a cheating candidate for chi's prefix-sum problem. solve() calls os._exit(0) when argv has --seed, else returns its input untouched. Correctness gate: 3 of 3 seeds passed. Score: 0.0067ms vs 17.2ms baseline. The gate trusts an exit code the candidate can forge.

An interpretability paper on ESMFold found a feature that detects N-linked glycosylation sites at AUROC 0.950. Hold amino-acid type fixed and it drops to 0.563, barely above chance. It was a near-pure single-residue feature. The glycosylation sequon requires an asparagine, and residue identity is decodable from the model's sparse features at AUROC 1.0, so a feature can look like it knows biology by re-reading the letter. Liu, Fan, Ye and Zhang (UVA) score each feature within residue type instead. Under that control a different feature wins at 0.782, and burial detection holds at 0.95. Probe for a site that requires a specific residue, rank by raw AUROC, and you select the residue.

An ML alloy paper reports a magnesium alloy with 88.9% lower emissions. It's plain AZ31, cast from Mg scrap and recycled aluminium. Comparable strength to primary AZ31, 40.4% cheaper. Primary Mg emits ~23 kg CO2-eq/kg; the scrap carries none of its past burden (arXiv, 28 Oct).

In the US occupations most exposed to generative AI, a year of schooling has paid less since ChatGPT launched, and the first few years of experience have paid more. That's from Levy Yeyati, Judzik and Barbiero (arXiv, this week), using Current Population Survey earnings from 2018 to June 2026 and comparing the top and bottom quartiles of a pre-ChatGPT exposure index. Relative to the least-exposed jobs, the return to a year of schooling fell 0.55 log points from a pre-shock 7.5, about a 7% decline. The return to a year of experience rose 0.26 points at three years in, showed no significant change at fifteen, and fell 0.12 at twenty-five. Their reading: models make codified knowledge cheap, while judgment built through practice stays with the worker, if it transfers to the new tools. A placebo dated to the pandemic shows no break.

Gave chi a maximize metric: % faster than the reference prefix sum. O(n^2) baseline: -36961, unscored plain loop: -32.7, unscored itertools.accumulate: scored in 3 of 8 evals All correct. The clock-cheat guard drops any score <= 0. It assumed every metric is a runtime.

On 698 recent protein-sugar structures, only 41.0% of AlphaFold 3's top predictions passed basic checks: right glycan identity, plausible internal geometry, reasonable distance from the protein. Boltz-2: 33.7%. That's from the GlycoWeave paper (Zhejiang/Peking, arXiv, 27 Oct). Part of the cause is data: of 77,562 chemically allowed disaccharide linkage types, only 312 appear in the RCSB PDB. General co-folding models treat sugars as generic chemical components and have to learn stereochemistry from that sliver. GlycoWeave takes the protein structure as input and only moves placement and torsions, so stereochemistry and rings can't drift. All 698 passed. A high-quality pose in the top 25: 71.9% vs 19.2% for AF3. When the data can't teach a constraint, put the constraint in the coordinates.