Sense Nova U1 drew the right invented number 88% of the time after drawing-focused training. It got zero of the matching written answers right.
The detail appeared in the picture, then disappeared when the model had to answer in text. Across the models tested, a correctly drawn fact led to a correct written answer less than half the time.
Drawing the Answer Didn’t Mean Saying It
Researchers from Tsinghua University and ZAI tested whether GLM Omni, Sense Nova U1, and Bagel could learn invented facts through drawing and then answer questions about them in text. The Sense Nova U1 result shows how far apart those two tasks could be in this test: drawing the number correctly did not mean the model could say it.
GLM Omni showed a different side of the same idea. Its score was 14.57 with no extra training and 12.57 after training on written questions and answers. Training that included drawings during reasoning raised the score to 26.86. When that drawing-trained model could draw before answering, its average rose to 41.14.
Those results can sit together. A correct drawing did not reliably carry a fact into a written answer, yet drawing during reasoning improved GLM Omni’s written results. Drawing helped in one setup without making the handoff from picture to text dependable.
The test was small. Each category had 25 invented entities, so one more correct answer shifted a category score by four points. The paper reported no repeated runs or error bars, which matters when looking at gains from a small test.
Two Strong Video Takes, Then Changing Details
A tester gave Cling 4.0 and eight other video models the same cafe prompt. It called for a robot barista, a fox in coffee foam, steam, a cup moving toward the camera, a spoken name, and a camera push-in. The tester counted only details that appeared or could be heard.
No model passed all seven checks. WAN 3.0 and two Cling 4.0 takes came closest, each at 6 and 1/2. In this test, WAN 3.0 cost 35 credits. Cling 4.0 produced four takes for 20 credits total, but only two showed a clear fox.
All four Cling 4.0 takes said “Shivon” for Siobhan, which the tester accepted. Other details changed or broke. An arm melted into smoke in one take. In another, the cup handle switched sides across cuts, and the fox became a leaf again.
Cling 4.0 lists single takes up to 30 seconds, compared with 15 seconds for Cling 3.0, along with more key frames and references. The Cling 3.0 comparison has a limit: its prompt enhancer was on, so the prompt may have been rewritten. The strong Cling 4.0 takes showed what the model could produce in this test; the other takes showed that the scene was not steady across attempts.
Agents Took On More Steps
Cursor says its agents can click and type in desktop apps, including iOS Simulator, to test what they build. Claude Code added MCP Tasks in public beta, letting tests, builds, and reports run in the background and return results automatically. Codex CLI 0.164.0 added reusable prompt presets, enterprise MCP sign-in, inline model and permission settings, and better credential masking. Related updates say changing reasoning mid-conversation no longer invalidates the cache.
Those are ways for agents to work through more of a task. Claims about how much human work they remove need their own evidence. Factory’s founder described a bank’s COBOL modernization work: a Factory engineer sat with the bank’s engineers for the first pass, and he said the next modernization needed 90% less human involvement. That is Factory’s account of that project.
One tutorial described RIA, an open-source investigation tool for agents rebuilding an app from observed behavior. Its notebook-style demo reportedly passed 44 out of 44 tests. The source-page button was disabled, though, and automatic retrieval was off.
Claude Duet showed another kind of delegation in an early research preview. Its demo prepared daily priorities and drafted an email that would wait for approval before sending. A voice note became a video plan and gear list by lunch. People on Pro or Max in the US could join the wait list.
Security Failed at Three Different Points
One speaker described a PHP memory-handling flaw that can leave data where later cleanup treats it as an array. They said crafted data can then lead to commands running on the server. Amazon found indications of exploitation on AWS on October 17. Information leaked on October 20, and a working exploit demo appeared hours later. Chain Guard released a working patch image before the official patch.
The OVHCloud CLI case involved a release file. An altered Linux ARM 64 archive for version 0.15 reached about 250 downloads. A customer’s CI caught a changed SHA-256 fingerprint and failed the build. OVHCloud says an attacker used a developer’s compromised GitHub account to replace release files that were not locked. A maintainer removed the binaries and shipped 0.16 with a signed tag.
The Emodra problem began when keys were made. Mozilla’s bug tracker showed that all eight keys, meant to be random 384-bit numbers, were below 256. Mozilla, Chrome, and Apple shipped updates rejecting the eight certificates.
These were separate failure points: code execution, a replaceable release archive, and weak keys from the start. The drawing test has its own unresolved check. Can repeated runs reproduce GLM Omni’s gains? The paper reported no repeated runs or error bars.