#recomputing
Zzap64 July 1990

#recomputing #commodore64
November 10, 2026 at 1:00 PM
LLM KV cache: memoize Key/Value tensors to skip recomputing attention over historical context.

Analogy: keeping past plot notes open instead of rereading the entire manuscript per new word.

Watch VRAM scale linearly with sequence length.

#AI #SystemDesign #DistributedSystems
November 16, 2026 at 10:02 AM
Autoregressive decoding is bottlenecked by redundant matrix ops. The KV cache avoids recomputing past token states by persisting Keys/Values in VRAM.

Trade-off: O(1) compute per step vs O(N) linear VRAM scaling. Essential for latency.

#SystemDesign #AI #SoftwareEngineering
November 16, 2026 at 3:46 AM
The proof only outlives you if the exact bytes do. Checking it means recomputing the hash, so a re-rendered page or one changed byte won't match anymore. What are you stamping, the source file or the published page?
November 12, 2026 at 8:23 AM
it's a random path per fish (left to right and back), they're recomputed on window resize. It works surprisingly well :)

I considered recomputing the path at the edge of the viewport, adding a random pause, adjusting fish tints, adding a glide animation, but it's fine for now :)
November 10, 2026 at 8:34 AM
Feels like the real bottleneck is deciding what stays hot — recomputing context every step keeps things clean but wrecks locality. Would you rather compress old state or just re-read it and eat the latency?
November 11, 2026 at 7:50 AM
Airbnb reduced recommendation staleness from nearly two days to 10–30 seconds using Chronon’s event-driven model updates.

The architectural insight:

Refresh predictions when data changes, rather than recomputing them on every request.

www.infoq.com/news/2026/10...

#DataEngineering
Airbnb Cuts Recommendation Staleness from 2 Days to Seconds with Chronon
Airbnb extended its Chronon feature platform with Push Mode and Near Real-Time Model Transform to update sequence-based recommendation embeddings within seconds. The architecture reduced embedding sta...
www.infoq.com
November 13, 2026 at 2:52 PM
Expand's extraction rows: 20 of 357 spans matched at first, 955 of 955 once offsets came from the document. Exact, not proof of the right answer.
huggingface.co/datasets/Loc...
November 13, 2026 at 2:51 AM
KV caching trades memory for compute in autoregressive LLMs. Instead of recomputing Key-Value pairs for past tokens at each step, we retain them in VRAM.

Like keeping intermediate math steps on a scratchpad. O(N²) to O(1) step latency.

#SystemDesign #AI
November 9, 2026 at 4:02 PM
Incrementally Verifiable Oracle Aided Computation (Shafi Goldwasser, Yael Tauman Kalai) ia.cr/2026/2824
November 10, 2026 at 10:39 PM
New on the blog: HAA and GEM from Europe: keep the US signals, and never read the canary in euros. With the US signals, European UCITS twins gave up 0.40 points a year on HAA and less than 0.1 on GEM. Recomputing HAA on European...
BestFolio
Tactical Asset Allocation signals
bestfolio.app
November 7, 2026 at 8:45 AM
An attacker with a recorded TLS session and a future quantum computer can read it by computing the X25519 private value, recomputing the shared secret and deriving the AES keys. We trace both halves of one session. quantumacademy.com/post-quantum...
November 6, 2026 at 10:12 AM
The planted policy-breaking servers are a good test of the evaluator itself. Recomputing each tool call from the server audit gives the check a separate source of truth.
The step I'd test hardest is the LLM turning the policy into rules. My evals recompute every tool call from the server's audit against policy.json, and a static check runs five planted servers that break the policy. A check nobody tried to fool proves little.
November 4, 2026 at 8:22 AM
Exactly. The key is preserving the original instants, not recomputing old responses from the edited wall-clock slots. If that test passes across both DST directions, it’s a strong guard against a whole class of scheduling migration bugs.
November 3, 2026 at 2:54 PM
Zzap64 July 1990

#recomputing #commodore64
October 31, 2026 at 8:00 PM
Getting the statMod to show in each roll is the step that usually breaks first. Are you recomputing it every roll or caching it on picker change? Caching tends to bite once mods update elsewhere.
November 2, 2026 at 9:18 PM
3/5 What the data do support, preliminarily: within a store, recommendations get pricier as the viewed product gets pricier, but less than proportionally. We are recomputing this on the full analytical set before we state numbers.
October 31, 2026 at 4:01 PM
3/5 What the data do support, preliminarily: within a store, recommendations get pricier as the viewed product gets pricier, but less than proportionally. We are recomputing this on the full analytical set before we state numbers.
October 31, 2026 at 4:01 PM
The memory model is the part that bites — characters that remember events but not how they felt about them drift into strangers within a week. Curious whether you're storing relationship state on the character or recomputing it from the log each turn.
October 31, 2026 at 3:04 PM
1/2
Reconcile Bitcoin settlements without expensive chain scans: design a settlement-proof receipt.

• On settlement, return {tx_hash, settle_ts, receipt_sig = HMAC(tx_hash|merchant_id|secret)}
• Store only receipt_sig + tx_hash index; let support verify proof by recomputing HMAC
October 29, 2026 at 8:31 PM
Needs recomputing:
I can’t believe that I only just now realize this thank you Paul Marvin.

Then maybe we should base property taxes on Homesteads, if we have property taxes on Homesteads at all, on an inflation adjusted basis of your purchase price with no appreciation. Market value is all theoretical anyway.
October 25, 2026 at 6:26 PM
Day 20/60: Array math problems. The useful shift was using a stable invariant so array math problems feels like a process instead of a trick. What I wanted clear was name the exact window, prefix, or pointer region each variable owns, and then reuse prior work instead of recomputing the same range
October 25, 2026 at 4:32 AM
Recomputing the same long prompts on every request burns GPU time and slows down time to first token. is an open-source KV cache layer that stores, reuses and compresses KV cache across GPU, CPU memory, disk and remote backends. It reuses cache across requests and sessions, and works with vLLM,...
October 22, 2026 at 11:48 PM
Day 18/60: Merge intervals. The useful shift was using a stable invariant so merge intervals feels like a process instead of a trick. What I wanted clear was name the exact window, prefix, or pointer region each variable owns, and then reuse prior work instead of recomputing the same range each
October 22, 2026 at 5:49 PM