FSR | Future Stack Reviews

@futurestackreviews.bsky.social

15 Followers 9 Following

$853/mo in SaaS. LLC in Osaka. I review what I pay for. If the review is wrong, I lose money before you do. →https://future-stack-reviews.com

Infographic showing the multi-query integration Fujitsu describes: one input problem, several query variants, multiple answers generated in parallel, results combined by majority voting or best-of-N, with the stated goal of more stable final answers. Four reading notes below, including that the 475× figure is a reported maximum experimental result rather than a promise for every deployment.
Infographic explaining Fujitsu PHOTON. Left panel: a standard Transformer processing text as a flat horizontal sequence of tokens. Right panel: PHOTON grouping tokens into semantic chunks and processing them hierarchically through stacked levels. Lower sections list the stated aims of lower memory pressure, higher concurrent capacity and potentially lower operating cost, followed by Fujitsu's reported figures: up to 475× multi-query throughput per GPU resource, model scales of 600M, 900M and 1.2B, and nine aggregated queries reported as sufficient to match a conventional Transformer.
Diagram showing four Kimi documents pointing at one self-serve pay-as-you-go API account. The API Help Center states no training or improvement. The OpenPlatform Terms allow broader content use unless agreed otherwise in writing. The Business Supplement gives no training for covered services. An order or checkout document may control the purchased service. A panel below reads that the public scope is not resolved.
Diagram of Gemini 3.6 Flash routed to three surfaces: the Gemini app where Future Stack Reviews ran its tests, the Gemini Developer API which is documented but untested, and the enterprise surface where two official Google pages disagree on Computer Use.
Four-panel comparison separating Gemini 3.6 Flash's publicly documented $7.50 output rate from vendor-reported benchmark results, each labelled with its source class and what it does not establish about production cost.
FSR framework. Zero code fixes is a status line, not a release decision. Keep a six-field scope record: path, scope, model evidence, findings, verification, disposition, and re-run when scope changes. Stop before real data: human security review and privacy scoping are required once an endpoint, personal data, or storage is added. Let a clean result feed a control, not close one.
FSR data panel. Security review: 20 agents, 14 findings (11 Info, 3 Low), consolidated to 10, zero code fixes, no safety verdict. Spec and quality audit: 16 agents, 10 candidates, 5 confirmed and fixed, 5 rejected as false positives. 36 of 36 agents logged claude-fable-5 with no fallback. One artifact, N=1. Independent blind re-review was not performed.
FSR scope map. The Fable 5 self-review covered one static index.html: markup, client-side scripts, and a lead form with no endpoint (form-action none), so no request left the page. The production surface was absent: endpoint, server validation, rate limits, storage and logging, sessions, HSTS and frame-ancestors. Zero fixes marked where the file ended, not that a production system was safe.
FSR pricing comparison verified June 18 2026. Top row reads: Real stack = Outreach (sending) + Credits (AI and data) + Healthy accounts (sending gate) + BYO APIs (extras). Three plan cards below: Outreach Growth $47 sends email but has no AI credits; Credits Growth $47 has 1,500 AI credits but cannot send; Starter Bundle $94 combines both. Bottom note: $9 / 150 credits equals about 30 agent-generated leads or replies.