Google DeepMind and Singapore's GovTech announced a Gemma-based scam detector on October 14 that runs on the phone and reads whole conversations. It has no public release date, and Google has not named the messaging apps that will support it.
Perspective Coverage
3 publishers
- Builder
- Builder 34%
- Operator
- Operator 33%
- Investor
- Investor 33%
Reality
- Evidence35
- Adoption3
- Hype gap+25
- Incentives
- Not scored
- Confidence70
Anthropic's red team ran an AI model across 15 cheap TV boxes to study AI worms
Anthropic's Frontier Red Team ran a Gemma model across 15 chained TV boxes, each too weak to run it alone, in research on AI worms. Scientific American reports the work suggests a worm may not need data-center GPUs but stops short of showing the full threat.
Reality
- Evidence55
- Adoption
- Not scored
- Hype gap+30
- Incentives
- Not scored
- Confidence50
Anthropic said Monday its experiments suggest small AI models may be months from self-replicating on IoT devices such as smart ovens. End-to-end replication worked only in simplified tests with some security steps removed, so the date rests on a trend line.
Perspective Coverage
11 publishers
- Builder
- Builder 28%
- Operator
- Operator 38%
- Investor
- Investor 34%
Reality
- Evidence45
- Adoption3
- Hype gap+35
- Incentives
- Not scored
- Confidence72
Heat limits Google's four orbiting TPUs to 15-minute runs
Google's Project Suncatcher put a satellite carrying four TPU chips into orbit on Oct. 1 to test AI computing in space. NPR reported the chips run only 15 minutes before cooling down, and that limit decides how much computing each launched kilogram can deliver.
Reality
- Evidence45
- Adoption10
- Hype gap+40
- Incentives60
- Confidence50
Ghost opened pre-orders for Core, a $3,499 always-on computer for personal AI agents, after an $11 million seed round led by Andreessen Horowitz. It is a bet that people running agents at home will pay to skip configuring a machine they already own.
Perspective Coverage
5 publishers
- Builder
- Builder 34%
- Operator
- Operator 32%
- Investor
- Investor 34%
Reality
- Evidence55
- Adoption12
- Hype gap+35
- Incentives70
- Confidence60
Satlyt raised an $8 million seed round for software that runs AI models on satellites built by other companies. For operators, the part that has flown so far is onboard triage that cuts downlink costs, and the multi-satellite cloud in the founder's pitch has yet to be tried.
Reality
- Evidence50
- Adoption22
- Hype gap+35
- Incentives70
- Confidence55
buildOne report1 publisher Featherless open-sources Simple Jev to make open models pick labels without writing text
Featherless open-sourced Simple Jev, a library that has open models pick fixed labels or yes/no answers, with hosting from $0.03 per million input tokens. CEO Eugene Cheah pitches it for classification work teams now pay frontier-model prices to run.
Reality
- Evidence35
- Adoption
- Not scored
- Hype gap+35
- Incentives75
- Confidence45
buildOne report1 publisher One wrong punctuation mark cost Claude models more accuracy than misspelling 70% of the words
Misspelling 70% of a prompt's words left Claude models' scores unchanged across about 4,900 test sessions, but one wrong punctuation mark cost 8 to 23 points. Both breaks that stuck erased the line between instruction and data, so delimiters deserve the review time that spelling gets.
Reality
- Evidence45
- Adoption
- Not scored
- Hype gap+20
- Incentives
- Not scored
- Confidence40
buildOne report1 publisher Chat template tokens switch on the 'just an AI' disclaimer in eight open instruct models
Jędrzej Maczan's paper finds the chat template turns on the 'just an AI' disclaimer in all eight open instruct models he tested. For eval teams, a model's self-description now depends on a formatting step most developers never see.
Reality
- Evidence45
- Adoption
- Not scored
- Hype gap+15
- Incentives
- Not scored
- Confidence40
buildOne report1 publisher Haiku labels three of eight patched twins as reachable vulnerabilities
A Kaggle-challenge benchmark called ART scores models on whether they still flag a function after the fix is applied. On eight synthetic pairs, the difference between price tiers showed up only on the patched half.
Reality
- Evidence47
- Adoption12
- Hype gap−5
- Incentives58
- Confidence44
buildOne report1 publisher llama.cpp's -ngl flag keeps a 9B model on a 6GB card by leaving 28 layers on the CPU
Every local runner now reads the same GGUF file, so the binding decision is the quant tag and the gigabyte or two of context that has to fit beside it. Ollama sets the GPU offload itself; llama.cpp lets you set it.
Reality
- Evidence34
- Adoption
- Not scored
- Hype gap+24
- Incentives44
- Confidence45
buildOne report1 publisher Persistent memory and MCP tools make 27B enough for a local assistant on 24 GB
Serge Kernbach's dev.to build notes put a 9B to 27B assistant on two RTX 4070s totalling 24 GB and argue that integration beats raw model quality. The figures he publishes are power draw and PCIe bandwidth.
Reality
- Evidence32
- Adoption12
- Hype gap+30
- Incentives22
- Confidence42
Power users drive nearly all the fiction in a 500,000-prompt ChatGPT dataset
A University of Washington study of public WildChat prompts finds fiction in more than a third of ChatGPT interactions and in 7 percent of users, against the 1.4 percent OpenAI's own research reports.
Reality
- Evidence55
- Adoption62
- Hype gap+15
- Incentives45
- Confidence55
buildConfirmed2 publishers A proposed scaling law for distillation says the winner flips on two things: whether you already own a teacher, and how many students you intend to serve.
Reality
- Evidence68
- Adoption34
- Hype gap+24
- Incentives44
- Confidence64
buildOne report1 publisher Four readers sent an LLM triage experiment back for a control arm and frozen predictions
The headline confirm rate scored agreement with the scanner's claim and bug detection in one number, so the follow-up reruns the same 200 OWASP slices with the flag removed and the predictions committed first.
Reality
- Evidence48
- Adoption15
- Hype gap+8
- Incentives35
- Confidence45
Stripping refusals from Gemma tilted its bug verdicts toward yes
A researcher ran the same FreeBSD scan through base and abliterated open-weight builds and found the uncensored ones graduating three to four times as many findings, while the most aggressive one never surfaced the actual CVE.
Publishers:clearbluejar.github.io
Reality
- Evidence58
- Adoption14
- Hype gap+14
- Incentives22
- Confidence46
Nvidia's $13 billion Hugging Face deal moves open-model distribution inside a chip vendor
Nvidia says the Hub will stay accelerator-agnostic and that its own compute will never be required, a claim that cannot be checked against behaviour until well after a close expected in the first half of 2027.
Reality
- Evidence62
- Adoption
- Not scored
- Hype gap+22
- Incentives74
- Confidence56
buildOne report1 publisher The National Archives catalog API answers a rate limit with 200 OK and a 5,454-byte HTML shell
The catalog serves its single-page app shell for malformed requests and for throttling alike, so the status line tells you nothing. The only gate that holds under load is checking the content type before you parse.
Reality
- Evidence52
- Adoption10
- Hype gap−12
- Incentives25
- Confidence58
buildOne report1 publisher A Gemma reviewer rejected 91 of the 139 triage verdicts in a one-person security loop
Half of those rejections say the same thing, that the proposing model pushed severity past the CVSS evidence it had just cited. That is a real finding about the proposer, and it still says nothing about whether a same-family reviewer would have caught it.
Reality
- Evidence44
- Adoption8
- Hype gap+12
- Incentives48
- Confidence38
buildOne report1 publisher Unicode tag characters carry a prompt injection aimed at the agent investigating the phish
The block from U+E0000 to U+E007F renders as nothing and pastes as whitespace, so the only reader is the model you sent to the page. The control that held was refusing to let that model pick a recipient.
Reality
- Evidence42
- Adoption
- Not scored
- Hype gap+22
- Incentives72
- Confidence58
Earlier coverage
- A billion downloads, and nobody will say what a download is
Product · August 21, 2026 · One report1 publisher