Alexander Whedon
@alex_whedon · 25.0K followers
Introducing SubQ - a major breakthrough in LLM intelligence. It is the first model built on a fully sub-quadratic sparse-attention architecture (SSA), And the first frontier model with a 12 million token context window which is: - 52x faster than FlashAttention at 1MM tokens -
SubQ is a real product by a real founder (@alex_whedon). RADAR has tracked 13.1M views on this launch, according to the source post linked below. RADAR measures how the launch reach was built, not whether the product works or whether anyone was honest. This reading is verified confidence and every input is public.
By Simba, Launch Intelligence Analyst · Reviewed by JK · Published 22 Aug 2026 · Confidence: verified
Independent, methodology-derived signal, not a statement of fact about any person. RADAR reads how reach was built, a signature, not an accusation. See the methodology.
The SubQ launch by @alex_whedon drew 13.1M views on 23.1K likes, which is 568 views per like, above the roughly 500 organic ceiling. RADAR reads a distribution-amplified (light) in how that reach was built, a signature of the mechanics and not a claim about the product or the founder. This is a verified reading and every input is public and reproducible.
New here? Start with the product
RADAR is FORKOFF's launch authenticity rating system. It reads whether a product launch earned its reach through real engagement or bought it through paid distribution, using only public signals anyone can pull from the launch post. Every reading carries a letter grade, a confidence label, and the date it was last checked, and links back to a published method you can reproduce. SubQ is a real product by a real founder (@alex_whedon). RADAR measures how the reach was built, not whether the product works or whether anyone was honest.
Distribution-amplified (light)
Independent, methodology-derived signal, not a statement of fact about any person. RADAR reads how reach was built, a signature, not an accusation. See the methodology.
What this grade means
RADAR reads a distribution-amplified (light) on this launch: the reach ran ahead of the engagement that organic reach produces. The grade describes how the reach was built, not whether the product works or whether anyone was honest. Paying for distribution is legal and common.
The signals RADAR reads
Views-to-likes ratio
Organic reach tops out near 500 views per like. When views climb far past that without the likes to match, the extra reach is arriving without the engagement organic reach produces.
Amplification wave shape
Organic amplification spreads over hours and days. A coordinated launch fires a synchronized burst of quote posts in the first few hours, read from each post's own timestamp.
Posting-time fingerprint
A post that fires exactly top of the hour on a weekday is scheduled. On its own it is weak, but it corroborates a coordinated launch alongside the other two signals.
Those three public signals sit on top of RADAR's five-component forensic read. The full method, the bands, and the confidence model are on the RADAR methodology page.
This launch in the data
Where it sits in the corpus
Rank 11 of 30 tracked launches by views per like, lowest (most organic) first. A lower ratio is the favorable end.
Against the benchmark
This launch's views per like next to the organic median (354) and the amplified median (1,441) across the tracked set.
The productSubQ sits in the foundation-model layer of the AI market, the small and heavily funded group of labs building the models everything else is built on top of. Its stated edge is architectural: standard transformer attention checks every token against every other token, an approach that gets quadratically more expensive as context grows. Subquadratic's pitch is that most of those comparisons do not matter, and a sparse-attention design that finds and processes only the ones that do can scale further for a fraction of the compute.
The launch post states three headline numbers: 52 times faster than FlashAttention at one million tokens, under five percent of the cost of Anthropic's Opus model at that context length, and roughly a thousand times less compute overall. These are the company's own stated benchmarks, cited here as the claim made at launch, not independently verified by RADAR. This reading is only about how the reach was built.
The launch was a single video-led announcement posted at 7:00 AM Pacific, on the hour, on Tuesday 5 May 2026, from Subquadratic co-founder Alexander Whedon's own account, @alex_whedon.
The post opens with a direct claim, "a major breakthrough in LLM intelligence," then states the two firsts back to back: the first model on a fully sub-quadratic sparse-attention architecture, and the first frontier model with a 12 million token context window. It follows with the two comparison numbers (52x faster than FlashAttention, under 5 percent of the cost of Opus) before closing on the compute-efficiency framing, "nearly 1,000x less compute and a new way for LLMs to scale."
The announcement ran from the co-founder's personal account rather than @subquadratic itself, a common pattern for an early-stage model lab where the founder's own following carries more real reach than a fresh company handle. Whedon's account was created in November 2024 and describes him simply as "Building better algorithms. Co-Founder at @subquadratic."
The launch leaned on a small set of assets. Read together they are a template a technical founder can repeat.
| Asset | What it did |
|---|---|
| Product video | The hook, demonstrating the model rather than only describing it |
| Two named benchmarks | FlashAttention speed and Opus cost, concrete comparisons a technical audience can check |
| A single superlative claim | First frontier model with a 12M-token window, the kind of specific, falsifiable claim that draws scrutiny and discussion |
| Timing: Tuesday 7:00 AM PT | Top of the hour, start of the US work week, the slot RADAR most often sees scheduled launches use |
Nothing here is unusual for a model-lab launch. A technical claim this specific invites both praise and pushback from an audience that can evaluate it, which is consistent with the heavy, two-sided written engagement the post drew.
The launch post reached 13,278,316 views on 22,511 likes, with 2,752 reposts, 1,486 replies, 1,962 quote posts, and 19,122 bookmarks. These are matured public metrics, re-pulled directly from the source post rather than a detection-time snapshot.
SubQ carries the largest raw like count of any launch in RADAR's tracked set at this reach, and that depth is the main reason the reading lands as a light lift rather than a heavy one. Under the 22,511 likes the post carried a combined 3,448 replies and quote posts, a heavy, costly written layer for a launch of this size. Replies and quotes are the hardest actions to manufacture because each is an original post a real person chose to write.
About 1.2 times the roughly 500 organic ceiling. The gap between reach and likes is small.
1,486 replies and 1,962 quote posts, the two hardest actions to fake at scale, both arrived in volume.
A retroactive sample of 40 quote-tweets and 39 replies found no tight-window engager clusters. No sign of a bought-engagement ring.
Zero sampled quote-tweets landed inside the first six hours; the wave built over the following days, the shape of word of mouth, not an injected spike.
A retroactive sample of the quote-tweets citing this launch found zero landing inside the first six hours, with the sampled wave building across the following three-plus days instead. That shape, a slow build rather than a front-loaded burst, is the pattern RADAR associates with word of mouth spreading through a technical community over time, not an injected spike timed to the launch window. The same sample found no coordinated engager clusters: 40 sampled quote-tweets and 39 sampled replies, checked for accounts posting within a two-minute window of each other, turned up none.
One number moved slightly since detection. The like count settled from 23,050 at first pull to 22,511 on this re-check, a small decline that is common as a platform prunes low-quality engagement over time; the view count rose from 13,090,265 to 13,278,316 over the same window. Neither shift changes the reading: the ratio moved from 568 to 590, still comfortably inside the light band, not the heavy band that begins at 2,000.
The SubQ launch shows what a technical, benchmark-led announcement looks like when it lands with the audience it is built for: a demonstration video, two comparisons a reader can independently check, one specific and falsifiable claim, and a launch slot at the start of the work week. The light amplification layer on top did not need to do much work, because the underlying claim was strong enough to generate its own written conversation.
FORKOFF builds launch videos made to travel and plans distribution around the claim a founder can actually defend. See how we approach it on the launch video service, or read the best product launch videos we track for more real examples.
Primary citation: x.com/alex_whedon/status/2051663268704636937. Every number traces to a public pull; reads re-checked over time.
Each named component carries a plain-English definition and a directional read where the public data supports one. RADAR publishes the component names, never the weights or the formula.
Whether the view curve grew the way organic spread does, or spiked like an injected burst.
Per-launch read not published in the public dataset. This component needs the forensic engine output.
Whether likes, replies, and reposts grew in step with views (the organic signature), or the views ran out ahead.
At 568 views per like, reach runs a step ahead of the likes: a light lift above the roughly 500 organic ceiling.
Whether the accounts replying are real, distributed people or a coordinated cluster posting together.
Per-launch read not published in the public dataset. This component needs the forensic engine output.
Whether the quote-tweet amplification looks like organic word of mouth or a known activation cluster.
Per-launch read not published in the public dataset. This component needs the forensic engine output.
Whether genuinely influential reference accounts engaged, or the reach was only low-quality volume.
Per-launch read not published in the public dataset. This component needs the forensic engine output.
Are you the founder of SubQ? You can claim or contest this read. RADAR attaches a founder response to the launch and re-examines any component you dispute.
Authorship
Simba
Co-founder, FORKOFF
Reviewed by: Kshitij JK
Last reviewed:
Published:
Methodology
RADAR verified reading of the SubQ launch from public metrics: the views-to-likes ratio against the roughly 500 organic ceiling and the posting-time slot, framed as a signature of how reach was built, not an accusation.
Sources cited
Where to go next
Three ways in, depending on what brought you here: learn how the score works, get a launch read or built, or get the plain answer on this launch.
Learn the score
What a launch authenticity score is
A launch authenticity score reads whether a launch earned its reach or bought it, from public signals. Start with the definitions and the checks you can run yourself.
Verify or build
Launch authenticity verification
Want a launch read by the same method, or a launch video made and distributed on the outcome? RADAR reads any public launch, and FORKOFF builds the launch behind the reach.
The skeptic's question
Is the SubQ launch legit?
If you are checking whether the SubQlaunch was real users or bots, here is the honest read: RADAR's reading is Distribution-amplified (light), at verified confidence, computed from public metrics and reproducible from the source post. It measures how the reach was built, not whether the product works.
Peer launches
Moda launch
@anvisha
4.5M views · 556 per likeRead
Parker launch
@alexgoughcooper
1.6M views · 498 per likeRead
Helena launch
@SeijinJung
3.7M views · 643 per likeRead
Cursor for iOS launch
@cursor_ai
6.4M views · 492 per likeRead
Durable launch
@jamesclift
3.6M views · 651 per likeRead
Contra Payments launch
@contraben
2.3M views · 445 per likeRead
The benchmark behind every reading
RADAR reads whether a launch's reach was earned or bought from public data, with the confidence label and the source citation on every reading.
AI SEO agency vs AI SEO software, decided with real Reddit AMA retainer figures, published tool pricing, and a stage-by-stage framework.
Read
9 real AEO agencies for AI and SaaS startups: real disclosed pricing where it exists, platform coverage, and the checklist that spots a relabeled SEO shop.
Read
Real disclosed AI SEO and AEO agency retainers for 2026, from $1,000/mo local deals to $50,000/mo enterprise deals, plus the tiers and red flags to watch.
Read
