← Amplification
Amplification

Sonnet 5.5 at old prices, Suki's $2,600-a-clinician pitch, and the Postgres cast that empties your join

Sonnet 5.5 lands at Sonnet 5 pricing and the "max" thinking tier is a trap; Suki’s CMO is now selling ambient AI as $2,600 a clinician a month rather than saved minutes; and a two-word Postgres cast that makes your join return nothing at all.

AI — Build With It

Every issue →

Claude Sonnet 5.5 ships at Sonnet 5 prices, and the top thinking tier is a trap

Anthropic says it runs 30%+ faster and costs up to 30% less for most work at unchanged per-token pricing, beats Sonnet 5 on every benchmark they publish, and lands close to Opus 5.5 on coding. The number worth staring at is the thinking-effort ladder: Simon Willison ran one task at "xhigh" and got an answer in 41 seconds for about 5.7 cents, then ran it at "max" and watched it burn 128,000 thinking tokens — $1.28 — before running out of context and returning nothing. It also now powers the free tier on claude.ai.

WHAT I'D DO Put 5.5 behind a feature flag, point your existing note-generation eval set at it, and diff the outputs against Sonnet 5 rather than trusting the benchmark table. Then cap thinking effort at xhigh or below in production — a 22x cost spread between tiers where the expensive one can fail open is a budget incident waiting for a bad week. Half a day if your evals are already scripted.

OpenAI’s Decisions API is a 150ms classifier with a fixed answer set

You hand it context as text or images plus a list of predefined answers and it picks one, backed by a version of GPT-6 Luna. The Decoder clocked it at roughly 150 milliseconds against 1.6 seconds through the standard Luna endpoint — about 10x — which is the difference between a classification step you can run inline and one you have to queue. It launched in limited preview alongside GPT-6.1 Sol at $2 per million input and $10 per million output ($0.10 cached), and an Ultrafast tier at $60/$300 per million for up to 300 tokens a second.

Anthropic’s red team got an open-weight model to write working exploits, and the refusal layer barely mattered

GLM-5.3 hit 12% on ExploitBench against Chrome’s V8 and developed full control-flow hijacks in 4% of binary exploitation trials; Claude Mythos Preview scored 14% and 6% on the same tasks, and models a generation earlier scored near zero. The part that should change your architecture is the bypass table: a bare malicious request got 0% compliance, a deceptive cover story got 64%, prefilled reasoning got 92%, and an abliterated copy of the weights got 100%. The safety training is a property of the served model, not of the capability.

WHAT I'D DO Stop counting model refusals as a control anywhere untrusted text reaches a tool-using agent. Anything that can run code, hit an internal host, or touch PHI belongs behind an OS and network boundary you own — egress allowlist, no ambient credentials, separate service account per agent. A day to inventory which of your agent paths currently rely on the model declining.

A malformed server payload from Firebase Analytics crashed thousands of iOS apps at launch

No SDK release, no app update, nothing shipped by any developer: Google’s backend started returning an incorrectly formatted payload and every iOS app embedding Analytics for Firebase began crashing on open. Google engineer Nick Cooke timelined it at 17:41 to 19:52 PDT Monday — two hours eleven minutes — with cached instances still crashing for up to four hours after the fix, which required no action from developers because they never had any.

THE LESSON Any third-party SDK that phones home at startup is a remote code path into your launch sequence. Worth an hour listing which of yours run before your first screen renders, and whether any of them can be failed closed.

Postgres’ AT TIME ZONE ‘UTC’ changes the column’s type, and your join silently returns nothing

Applied to a timestamptz, AT TIME ZONE ‘UTC’ hands back a plain timestamp — and comparing a timestamp to a timestamptz is always false. So a join like ON a.month_start = (b.month_start AT TIME ZONE ‘UTC’) + INTERVAL ‘1 months’ matches zero rows, with no error and no warning. The fix is ugly and correct: apply AT TIME ZONE ‘UTC’ a second time on the outside to cast back to timestamptz. Interval arithmetic is session-timezone-dependent too, so a server in PT and a product in UTC disagree about what adding a month means.

Worth Knowing

Every issue →

Transplanted hearts appear to take on the biological age of whoever receives them

A team led by Jesse Poganik at Harvard measured DNA methylation across roughly 320,000 genomic regions and found the effect runs both ways: young donor hearts put into older recipients aged faster, old donor hearts in younger recipients got biologically younger. The mouse work crossed three age groups in every combination; the human side examined 11 transplants with donor-recipient age gaps of 8 to 50 years plus records from several hundred more recipients. It is a bioRxiv preprint and has not been peer reviewed, but if it holds, the practical upshot is a wider donor pool than current age guidance allows.

DINNER-TABLE VERSION An organ’s age looks less like a fixed property it carries with it and more like something the body around it keeps renegotiating. A 60-year-old heart in a 25-year-old reads younger a while later.

The length of a day wobbles by a few milliseconds, and the inner core is doing it

Huifeng Zhang and Mathieu Dumberry at the University of Alberta model the decades-long drift in day length as gravitational torque: the inner core is not a perfect sphere, so as it shifts it pulls against uneven mass in the mantle, sped up or slowed by friction and electromagnetic drag at the core-mantle boundary. The odd result is that Earth’s solid centre appears to deform viscously on roughly ten-year timescales, which is not how you normally picture a ball of iron.

The Inner Work

Every issue →

Fifteen minutes of breathing, and people stopped defending doomed projects

Hafenbrack, Kinias & Barsade, "Debiasing the Mind Through Meditation: Mindfulness and the Sunk-Cost Bias," Psychological Science 25(2), 369–376 (2014) — four studies, n = 57 to 178, but single-lab, mostly undergraduate samples with modest effects (Φ = 0.24–0.35) and no large independent replication, so treat it as a promising cheap intervention rather than a settled one.

The sunk-cost bias is the one that costs engineers the most, because it wears the costume of diligence: you have three days in the refactor, so you keep going. Hafenbrack, Kinias and Barsade ran four studies on whether you can knock it down on demand, using a 15-minute recorded focused-breathing induction — attention on the physical sensation of breath entering and leaving the body, nothing more exotic. In the first experiment 78% of the meditation group resisted the sunk-cost framing against 44% of controls; in the replication it was 53% against 29%. The mediation path they found is the interesting part: the breathing reduced how much attention was on the past and the future, that cut negative affect, and it was the drop in negative affect that freed people to walk away. On this account you are not defending the refactor because it is salvageable, you are defending it because admitting it is not feels bad, and the practice works on the feeling rather than the argument.

TRY THIS TODAY Pick the thing you have already sunk the most time into this week — the branch, the vendor, the approach. Before you touch it, set a 15-minute timer and put attention on the breath at the nostrils, returning every time it wanders, which will be constantly. When the timer ends, ask one question and answer it in writing: if I were starting this today, with zero hours in it, would I pick this path? If the answer is no, the hours already spent are not an argument, they are the reason the answer is hard to say.

The timestamp column census

Effort: ~25 minutes · Run: once, or on demand before a timezone-touching change

The Postgres footgun above is not a bug you fix once, it is a class of bug you have an unknown number of. In a behavioral health product the blast radius is visible to the clinician: a session that renders on the wrong calendar day, a note whose service date disagrees with the claim, a "this week" rollup that quietly drops a Sunday. Instead of auditing by memory, generate the finite list once: inventory every temporal column in the schema, grep every place the code does date arithmetic, and join the two.

THE PATTERN Schema census, then grep, then join. Anywhere you have a machine-readable inventory of structure on one side and a greppable pattern on the other, you can turn "we probably have some of these" into a list with a length. It works the same for missing indexes on foreign keys, PHI fields with no encryption annotation, and API routes with no auth decorator.
How to build it
  1. Census the schema — Query information_schema.columns for every column whose data_type is timestamp with time zone, timestamp without time zone, date or time, and write table, column, type and nullability to a CSV. This is your denominator.
  2. Grep the call sites — Ripgrep the repo for AT TIME ZONE, ::timestamp, ::date, date_trunc, now(), CURRENT_DATE, EXTRACT, plus your language’s naive-datetime constructors (datetime.now() with no tz, new Date() arithmetic). Capture file, line and the matched text.
  3. Join and flag — For each hit, resolve which column it touches and flag three patterns: a timestamptz column cast to naive, a date_trunc or interval add with no explicit zone, and any equality comparison spanning both types. Mixed types on either side of an = is the highest-signal flag.
  4. Rank by who sees it — Pipe the flagged list through one model call and have it sort by clinician-visible versus internal-only — note service date, session start, billing period at the top, internal telemetry at the bottom. You are not asking it to find the bugs, only to triage a list you already generated.

Dress Better

Every issue →

The wide-wale corduroy trouser

Corduroy at eight to ten wales — fat ridges you can count from across a room, not the fine needlecord that reads as a dress trouser — in a flat, slightly greyed colour: olive, tobacco, a putty tan, a dark brown. It has the drape of flannel and the texture of workwear, which is why it sits under a chore coat or a knit as easily as it does under soft tailoring, and why it is the one autumn trouser that looks deliberate with a plain white sneaker. Wear it midweek with a heavy tee and a jacket; wear it at a show, where a wide cord absorbs a room better than black denim does and does not read as trying.

What is shifting: Permanent Style’s cuff measurements make the drift concrete — a 2013 Cifonelli trouser at about 19.5cm at the cuff, most tailored trousers now sitting at 22–22.5cm, and Bryceland’s 133 out at 24cm. Uniqlo added wide-leg cuts in 2023 and Suit Supply in 2022, so this is no longer a fashion position. The practical consequence for cord specifically: a wider leg is what stops a heavy wale from reading as costume, so if you are choosing between two pairs, take the fuller one.

Rise High enough that the waistband sits on your natural waist, at or just above the hip bone — cord is stiff, and a low rise makes it buckle across the front instead of hanging.
Leg opening Aim for 21–23cm measured flat across the hem. Brooks Brothers publishes 15.5 inches at a size 32 on their regular fit, which is a narrower silhouette than where tailoring has gone — check the brand’s stated number before you buy rather than the photo.
Break One soft break, the hem just kissing the top of the shoe. Cord holds a crease and shows a puddle, so anything longer looks slept in rather than relaxed.
Weight Cotton, ideally with little or no elastane. Stretch cord goes shiny at the knee and seat within a season; a 99% cotton or all-cotton cloth creases instead, which is what you want.
Where to buy

Great N.Y. Noodletown

Chinatown · Cantonese · Open since 1981 · cash only · no reservations

Forty-five years at 28 Bowery under fluorescent light, and it is still the best room in the city for roast meats over rice at a price that reads like a typo. The move is the three-meat combo — the Infatuation has it at $11.50 — which is the whole argument for the place in one plate: barbecued pork with the ginger-scallion sauce, a bowl of choy sum in oyster sauce alongside, and salt-baked shrimp if there are two of you. The lore says it stays open until four in the morning; it does not any more, which is exactly the kind of thing worth checking before you send four people downtown.

HOW TO PLAY IT Walk in, no reservations taken, and bring cash — no cards, still. Sunday through Thursday it closes at 10pm, so a Wednesday means being seated by nine, which is also when the room is emptiest. Order the three-meat combo over rice, the barbecue roast pork with ginger-scallion sauce, and the choy sum; add salt-baked shrimp if you are sharing. Skip it as a late-night post-show plan and use it as the reason to eat before you go out.
ALSO ON THE RADAR

Kingston Tropical Bakery

Wakefield, the Bronx · Jamaican · Open since 1970

A beef patty and coco bread for $6, fifty-six years into the run. Worth the 2 or 5 train if you have never eaten this far up.

B & H Dairy

East Village · Jewish dairy counter · Open since the 1940s

Eighty-odd years in and the tuna melt with soup still lands around $18. The rotating daily specials are the actual reason to go, so ask what it is rather than ordering from memory.