🔍 Read the full analysis: Jev Can See Now: My Experiment With 30 Personal Pictures on ThorstenMeyerAI.com
Get hardware and tech essentials delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Thorsten Meyer tested TypeSafe’s new image support in Jev on 30 featured images from his website, comparing Jev 1.14 with 1.13. In the reported product-photo tests, Jev 1.14 matched each image to its own headline more strongly than to unrelated headlines and identified 9 of 10 products correctly. The source cuts off during its account of news-image results, so the full findings cannot be established.
Thorsten Meyer tested TypeSafe’s new image support for Jev on 30 featured images from his website, reporting strong results for product photos but an incomplete account of how the model handled news images. The experiment, published October 15, followed the October 14 release of Jev 1.14, which lets users include images in decision-model requests and adds a bounding-box response type.
Meyer used the site’s 10 product roundups, 15 news posts and five essays. For each image, he asked whether it matched its own headline, a randomly selected other post’s headline and an unrelated headline about repotting a houseplant. He also asked Jev to identify which of 10 products appeared in each product image. He says he ran the tests on both Jev 1.14.0 and Jev 1.13.0, using about 60 requests in total.
For product images, Meyer reports that Jev 1.14 assigned each image a higher fit score for its own headline than for a randomly chosen headline. Own-headline scores ranged from 0.65 to 0.89, with a median of 0.76; the random-headline median was 0.06, and no score for the unrelated plant headline exceeded 0.10. Jev 1.14 also named the correct product in nine of 10 images.
The one product-identification miss involved a photo showing a router with antennas beside a cylindrical mesh unit. Jev called it a Wi-Fi 7 router, with confidence of 0.60. Meyer says the correct identifications had confidence scores of at least 0.77. Jev 1.13 identified all 10 products correctly, but two answers had confidence below 0.5. These are results from Meyer’s small, self-selected test, not an independent benchmark.
Jev 1.14 · images + box
Jev can see now. Here is what 30 of my images showed.
I asked about every featured image three times: does it fit its own headline, a random other post’s headline, an off-topic headline? Both Jev versions, about 60 requests.
Does this image fit the headline?
Median “yes” probability on Jev 1.14, by kind of post
- Product roundups10 images0.760.061.13: 0.79 / 0.05
- News posts15 images0.330.161.13: 0.51 / 0.25
- My essays5 images0.510.161.13: 0.62 / 0.33
Off-topic headline (“How to Repot a Monstera Plant”): never above 0.10 on either version, 60 of 60 rejected.
Why news is weak: almost all news images show the same scene, people at desks looking at monitors. Jev could not tell them apart, and it was right not to.
1.14 is stricter than 1.13
Same 30 images, same questions
| Result | 1.14 | 1.13 |
|---|---|---|
| Fits own headline (≥ 0.5) | 15/30 | 23/30 |
| Fits a wrong headline (≥ 0.5) | 1/30 | 4/30 |
| Own beats random other (AUC) | 0.91 | 0.89 |
| Names the product (10 options) | 9/10 | 10/10 |
1.14’s one product miss: mesh system read as “Wi-Fi 7 router”, confidence 0.60. Re-tune thresholds before jev-latest moves.
What shipped on Oct 14
- headerImagesin state, instructions or criteria · 1.13 + 1.14 ·
TypeSafe-Danger-Zone-Images - 1.14 onlyBox typeone bounding box + confidence. 6 of 6 clear targets boxed (0.66–0.90); an absent bicycle still got a box (0.03)
- headerLong context64k / 32k → 262k tokens, slower and less accurate
- opt-inHelp find weaknessesqueries used for testing, not training
now: jev-preview = 1.14~Oct 21: jev-latest → 1.14notice: 60 days (1.13), 30 days (1.14)
Source: own test on October 15, 2026. 30 featured images from ThorstenMeyerAI.com (10 product roundups, 15 news posts, 5 essays), JPEG at up to 1,568 px, jev-1.14.0 and jev-1.13.0 with the image header. Bars show medians. AUC = probability that an image’s own headline scores higher than a random other headline. Features, prices and limits: TypeSafe blog post “Jev 1.14” and docs.typesafe.ai, as of October 15, 2026.
Powered by Thorsten Meyer AI
What the Product-Photo Test Shows
The experiment offers an early, practical example of what image-based classification and scoring could do in a publishing workflow: check whether a featured image fits a headline or identify a product in a photo. Meyer describes that check as one that often goes undone by hand. If repeated testing supports the result, such a system could help flag mismatches for editorial review, rather than replace that review.
The findings also show why confidence scores and human checks matter. Jev 1.14 made one product mistake, while the earlier version got the product labels right but gave two answers low confidence. The test does not establish how the model performs across other publishers, image types or larger datasets. It is an individual experiment, and the reported numbers should be read on that basis.
As an affiliate, we earn on qualifying purchases.
Jev Adds Images Behind Opt-In Headers
Before this release, Jev’s decision questions were limited to text, according to Meyer. TypeSafe’s October 14 post introduced image inputs for Jev 1.13.0 and 1.14.0, as well as a new box response type in 1.14 that returns one bounding box for a described target. Image support is activated with the request header TypeSafe-Danger-Zone-Images: true; normal production behavior remains unchanged without the opt-in.
TypeSafe said the features remain too unreliable for it to endorse production use, and placed them behind what it calls a danger-zone header. The release also adds an optional longer context limit, from 64,000 tokens per request and 32,000 per question to 262,000 for each, though the company advises against the larger limit because reliability can fall on long inputs. Jev 1.14 is available as jev-preview; TypeSafe said jev-latest would move to it about a week after release.
Images are sent through an asset map and referenced in the request text, Meyer explains. The supported still-image formats are PNG and JPEG, with size and request limits; WebP and HEIC files need conversion. TypeSafe says there is no separate image surcharge: image input is billed at the model’s ordinary input-token rate. The release also includes an opt-in that lets TypeSafe use queries to identify model weaknesses, while the company says it will not train on those queries.
“There is “still way too much jaggedness for us to vouch for these for production use cases.””
— TypeSafe, as quoted in Meyer’s report
product photo identification tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
News-Image Findings Are Incomplete
The supplied source material ends partway through the section about news images, after the sentence “The news posts were a dif”. It does not provide Meyer’s results for the 15 news images, the five essays, or a complete comparison between Jev versions across the full test. Those findings cannot be inferred from the product-photo results.
The test also does not report an independent replication, a larger evaluation set or a formal benchmark. TypeSafe has cautioned that the experimental features are not ready for production use. It is unclear from the provided material whether Meyer plans to repeat the test or publish additional results.
As an affiliate, we earn on qualifying purchases.
Release Schedule and Further Testing
TypeSafe’s stated schedule is for jev-latest to switch to Jev 1.14 about one week after the October 14 release, around October 21, 2026. The company has described deprecation notice periods of 60 days for Jev 1.13.0 and 30 days for Jev 1.14.0, with a formal policy still to come.
For the image features, TypeSafe’s position is that users should opt in and assess performance for their own applications rather than assume production reliability. Meyer’s report, as supplied, does not include the rest of the news-image findings. Further evidence would be needed to judge how consistently Jev handles varied images and headline types beyond this 30-image test.
photo bounding box annotation software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did Meyer test?
He tested whether 30 featured images from ThorstenMeyerAI.com matched their headlines, alongside unrelated headline checks. For 10 product photos, he also asked Jev to identify the product.
How did Jev 1.14 perform on the product photos?
According to Meyer, Jev 1.14 identified the correct product in nine of 10 images. The reported own-headline fit scores ranged from 0.65 to 0.89, with a median of 0.76.
Was Jev 1.14 shown to be more accurate than Jev 1.13?
Not overall. Jev 1.14 got nine of 10 product identifications correct, while Jev 1.13 got all 10 correct in Meyer’s test. The older version gave two of its correct answers confidence scores below 0.5. The source does not provide complete results across all image categories.
Can the new image features be used in production?
TypeSafe says the features are not yet reliable enough for it to vouch for production use. Image support requires an explicit opt-in header, and users should evaluate it for their own needs.
What is still unknown about the experiment?
The supplied report cuts off during the news-image section, so its results for news posts and essays are unavailable. The test also does not establish performance across larger or independently selected datasets.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
