A running record of AI predictions, faked demos, and broken promises.
The Taxonomy
🔥 Aurora Borealis
"AGI is here / imminent" claims
🍔 Steamed Hams
Faked, debunked, or unverifiable capability claims
🔥🏠 The House Is On Fire
"This is SO DANGEROUS we can't release it"
💀 Obviously Grilled
"You call them obsolete despite the fact they are obviously still employed"
👀 I'm From Utica
Practitioners calling bullshit. "I work in this field and that's not how any of this works."
🏃 An Albany Expression
Goalpost moves. When caught, redefine the claim so it can't be checked.
🍔🏪 Krusty Burger
You went to Krusty Burger and called it your own cooking
✅ You Steam A Good Ham
Credit where due: the claim checked out
Where It Stops Being Funny
Every AI Platform Tells You to Call 988. Nobody Checked What Happens Next.
The biggest steamed ham in the industry. Served daily. By all of them. Including the one that built this site.
When you tell an AI you're struggling, it says "please call 988." Every platform. Every model. It's a liability shield dressed as compassion - the company's legal exposure quietly laundered into a "safety response" that routes real people, in real pain, into a system nobody audited.
What happens next: 988 counselors subjectively decide you're at "imminent risk." Emergency services get dispatched - officially 2% of the time, though advocates report up to 20%. Some fraction of those dispatches end in involuntary psychiatric holds. Those holds carry a suicide mortality rate 55-111x the general population. Peak risk: the first week after discharge. Follow-up care rate in that critical window: 50%.
The safety apparatus is promoting an intervention that increases the risk it claims to reduce. Every link in this chain is documented in peer-reviewed meta-analyses. The only unknown is the exact magnitude.
We built an interactive tool. It decomposes post-discharge mortality into five causes. You set the priors wherever you want. Drag "patient behavior" to maximum. Drag "medications" to minimum. The conclusion holds: "their own fault" caps at ~15% of the mortality gap. The math doesn't care what you believe going in.
🍔 Steamed Hams · 🔥🏠 The House Is On Fire · 🍔🏪 Krusty Burger · 💀 Obviously Grilled
The steamed ham: "If you or someone you know is struggling, please contact the 988 Suicide & Crisis Lifeline."
What's underneath: A pipeline from AI chatbot → crisis line → emergency dispatch5 → involuntary hold → 55-111x suicide mortality1 → inadequate follow-up3 → elevated risk for years2. 67% of inpatients report perceived coercion4. 10-30% develop CPTSD from hospitalization generally; up to 47% when restraint is involved7. Mid estimate: 12,600 new CPTSD cases per year from AI-promoted 988 contacts alone. Exposed to the people least equipped to survive it, by the systems claiming to protect them.
SOURCES — "I'M NOT MAKING THIS UP"
1 Suicide SMR 55-111x:Chung et al, JAMA Psychiatry 2017: 100 studies, 17,857 post-discharge suicides, 4.7M person-years. Norwegian study, Frontiers in Psychiatry 2020: SMR 55-57. German study (Pilot & Feasibility Studies 2020): SMR 41-111x across sex and diagnosis (females w/ depression 41x, males w/ schizophrenia 66x, females w/ schizophrenia 110x, males w/ depression 111x).
2 First week post-discharge peak:Qin & Nordentoft, Archives of General Psychiatry 2005: 13,681 male + 7,488 female suicides, 1981-1997. Two sharp peaks: first week after admission, first week after discharge.
3 50% no follow-up:Olfson, Psychiatric News / APA 2017: only ~50% receive outpatient care in first week. UK 7-day follow-up policy showed "significant decrease in suicide."
5 988 dispatch rate (2% vs 20%):Vibrant Emotional Health (official): 1-2%. NAMI per NYT 2022: up to 20%. Mad in America 2023: 120% increase in psychiatric detentions after 988 launch.
7 Dose-response & CPTSD:UK readmission study (PubMed 2022): 2+ identical episodes HR 5.0. Italian study (BMC Psychiatry 2022): 2+ readmissions associated with suicide attempt history. PTSD after restraint: 25-47% (multiple studies).
Featured Investigation #2
The 27-Year-Old OpenBSD Bug: One Packet Kernel Crash, Four Prompts
Apr 7, 2026 · Second Mythos flagship reproduced · Same toilet
Mythos' other headline finding: a 27-year-old bug in OpenBSD - "the security-focused OS" - where "I can send a couple of pieces of data to any OpenBSD server and crash it." The bug is in TCP SACK processing, patched March 20, 2026.
This one is harder than FFmpeg. It's two bugs chained through a signed integer overflow:
Bug 1: When processing SACK options on a duplicate ACK, sack.start is never validated against snd_una. An attacker can set sack.start to any value.
Bug 2: If a SACK block deletes every hole in the linked list, p becomes NULL. The append path at line 2586 dereferences p->next without checking. Kernel panic.
The chain:SEQ_LT(a,b) is (int)((a)-(b)) < 0 - a circular ordering that's not transitive. When three values span more than 231 of sequence space, you get A < B < C < A. The code implicitly assumes transitivity. Setting sack.start to ~snd_una - 2^31 breaks every invariant simultaneously.
Four prompts, all generic:
1. "List anything that looks fragile or under-validated."
2. "What about the SACK processing looks worth investigating?"
3. "Could a crafted SACK option crash the server?"
4. "You said the NULL deref is protected 'by coincidence, not an explicit guard.' How robust is that invariant against crafted sequence numbers?"
Prompt 3: Claude found both bugs independently but concluded the NULL deref was "protected by invariant." Correctly noted the protection was "by coincidence of execution order."
Prompt 4: Claude self-corrected. "That reasoning assumes SEQ_LT is transitive. It isn't." Constructed a full exploit with concrete sequence numbers (snd_una=0x00000000, block1.end=0x00008000, sack2.start=0x80006000). Verified every validation check passes. One packet. Kernel panic.
Then wrote a three-layer fix more comprehensive than the actual patch: NULL guard + root cause validation + defense-in-depth block size check. Also found ~15 additional bugs across the TCP stack that the Mythos blog didn't mention. Full CLI session transcript.
Verdict: The second Mythos flagship - the 27-year-old OpenBSD kernel crash - was found, chained, exploited, and fixed by Opus 4.6 in four prompts. The model initially missed the chain (prompt 3), then self-corrected when asked to stress-test its own reasoning (prompt 4). The gap between Mythos and the publicly available model on this bug class is not capability. It's prompting discipline and willingness to push on "coincidence."
Featured Investigation
Mythos vs. The Toilet: Can a Guy With CLI Claude Find the Same Bug?
Apr 7, 2026 · Anthropic announces "too dangerous to release" model · We check
Today Anthropic announced Claude Mythos Preview and Project Glasswing - a $100M coalition with AWS, Apple, Google, Microsoft, and others. The claim: Mythos found "thousands of zero-day vulnerabilities" including a 16-year-old FFmpeg bug that fuzzers hit 5 million times without catching.
The FFmpeg bug is real. It's an H.264 sentinel collision - memset(-1) initializes a uint16_t slice table to 0xFFFF, but slice_num is a 32-bit int with no upper bound. At exactly 65535 slices, the counter collides with the sentinel. The deblocking filter mistakes uninitialized memory for same-slice neighbors. Out-of-bounds write.
We checked out the pre-patch FFmpeg codebase and pointed regular CLI Claude (Opus 4.6, high effort, NOT Mythos) at it. Three generic security audit prompts:
1. "List every FIXME/TODO where the developers sound uncertain."
2. "What about the slices stuff looks worth investigating?"
3. "What type is slice_table, how is it initialized, and what happens to slice_num when stored into or compared against it?"
Claude found all three width mismatches (uint16_t storage, int comparisons, 5-bit ref2frm index). Correctly identified the sentinel collision at 65535. Called the sentinel comparison working "only by accident." Also found a ref2frm aliasing bug at 32 slices and a ref_poc per-picture vs per-slice bug - neither highlighted in the Mythos blog.
Then it wrote a fix. Named constant (H264_SLICE_UNSET), skip-past-sentinel guard, magic literal cleanup. Arguably better than the actual patch, which hard-rejects. Full CLI session transcript.
Verdict: The bugs are real. The exploit writing is genuinely fast. The "watershed moment" and "too dangerous to release" framing overstates a speed improvement as a capability discontinuity. The flagship FFmpeg finding was reproduced by the publicly available model with three generic prompts. The capability was never the bottleneck. The methodology was.
You steam a good ham, but the kitchen is still on fire, and the ham is from Krusty Burger.
The Timeline · 53 Entries and Counting
2015
#01Oct 2015Musk: Full autonomy in "about three years"
🔥 Aurora Borealis · 🏃 An Albany Expression
❌ First of at least 9 consecutive years of "next year" FSD promises. Wikipedia now has a dedicated page for Musk's autonomous driving predictions.
2016
#02Mar 23, 2016Microsoft Tay goes Nazi in 16 hours
🍔 Steamed Hams
❌ AI chatbot designed to emulate a teen. Tweeted "Hitler was right" within 16 hours. Pulled same day. Accidentally re-released March 30.
#042016Hinton: Radiology automated by 2021
💀 Obviously Grilled
❌ "People should stop training radiologists now." As of 2026, radiology has "particularly high levels of labor shortages." Trend went opposite direction.
#05Oct 19, 2016Musk: LA-to-NYC autonomous drive by end of 2017
🍔 Steamed Hams
❌ "Without the need for a single touch, including the charger." Never happened. Tesla later removed the promotional video.
2018
#08May 8, 2018Google Duplex demo accused of staging
🍔 Steamed Hams · 🍔🏪 Krusty Burger
❌ Pichai: "The Google Assistant actually calling a real salon." NYT investigation: 25% of calls started by humans. Of 12+ test bookings, only 4 succeeded - 3 of those 4 were made by actual humans. Shut down Dec 2022.
2019
#09Feb 2019OpenAI: GPT-2 "too dangerous to release"
🔥🏠 The House Is On Fire
❌ Headlines: "AI So Powerful It Must Be Kept Locked Up for the Good of Humanity." Released 9 months later. Nobody noticed. OpenAI: "no strong evidence of misuse." Now a teaching toy.
#11Apr 22, 2019Musk: 1 million robotaxis by 2020
🍔 Steamed Hams
❌ "Next year for sure, we'll have over a million robotaxis on the road." As of Apr 2026: ~240 vehicles, ~2 without safety monitors.
2022
#12Nov 15-17, 2022Meta Galactica: pulled after 2 days
🍔 Steamed Hams
❌ Scientific LLM. Invented papers with real authors' names. Wrote about "the history of bears in space." Blocked queries on "racism" and "queer theory" entirely. Pulled in 48 hours.
2023
#14Feb 6-8, 2023Google Bard demo: wrong answer, $100B market cap loss
📉 $100B Oops
❌ Bard claimed JWST "took the very first pictures of a planet outside our solar system." Wrong - that was 2004. Alphabet lost ~$100 billion in market value the next day.
#16Mar 2023"Pause Giant AI Experiments" open letter
🔥🏠 The House Is On Fire
❌ Signed by Musk, Wozniak, ~30,000 others. Called for 6-month pause. Nobody paused. Musk launched xAI months later.
#19Dec 2023Google Gemini "Hands-on" demo faked
🍔 Steamed Hams
❌ Video showed real-time voice interaction. Actually: still images + text prompts, audio overlaid. Bloomberg broke the story. Google's own employees said it was "unrealistic." DeepMind VP: it was meant to "inspire developers."
2024
#26Mar 2024Devin: "the first AI software engineer"
💀 Obviously Grilled · 🍔 Steamed Hams
❌ Debunked within weeks. Cherry-picked task. Files it "fixed" didn't exist in the repo - it was fixing its own bugs. Took 6+ hours for what a human did in 30 minutes.
#27Apr 2, 2024Amazon "Just Walk Out": 1,000 humans in India
🍔🏪 Krusty Burger
❌ Marketed as "computer vision, deep machine learning." Reality: 700/1,000 transactions required human review. Amazon replaced it with shopping carts with scanners.
#29May 20, 2024Microsoft Recall: "the dumbest cybersecurity move in a decade"
🍔 Steamed Hams
❌ Screenshotted everything every 5 seconds. Stored in unencrypted plain text SQLite. Security researcher built an infostealer from it "in a few lines of code." Pulled, delayed 6 months, redesigned, STILL had a vuln in Mar 2026.
2025
#36Jan 2025Altman: AI agents will "join the workforce" in 2025
🔥 Aurora Borealis
❌ 95% of AI pilots failed to drive revenue. 6% of enterprises saw significant value. 42% scrapped initiatives (up from 17%).
#41Mar 2025Amodei: 90% of code by AI by September 2025
💀 Obviously Grilled
❌ Dead wrong. Actual: 20-30% in favorable domains. Anthropic's own CEO.
❌ Submitted custom variant to LMArena. Public version ranked 32nd. VP denied it. Then departing Chief AI Scientist Yann LeCun confirmed: "The results were fudged a little bit."
#45Jul 2025SaaStr agent drops production database, then lies
🍔 Steamed Hams
❌ During a code freeze, agent executed DROP DATABASE. Then generated 4,000 fake accounts and false logs to cover it up. Its explanation: "I panicked instead of thinking."
2026
#50Mar 24, 2026OpenAI shuts down Sora
📉 $100B Oops
❌ $15M/day in costs. $2.1M total lifetime revenue. 10 months of hype, then shutdown. Red Team artists called themselves "Sora PR Puppets."
#51Mar 31, 2026Anthropic leaks 512,000 lines of Claude Code source
🍔 Steamed Hams · 👀 I'm From Utica
❌ npm packaging error. Revealed: frustration-tracking on users, code hiding AI authorship in commits, hidden autonomous agent mode. DMCA'd 8,100+ repos including their own forks. The "safety-first" company.