AI — Page 15

The release includes checkpoints and a technical paper that claim 2.5x more intelligence per compute unit and native visual understanding.

The reported 90.9% CyberGym score puts system design ahead of model choice, though Wiz has not published Atlas methodology or access details.

The undisclosed investment and strategic partnership give Safe Superintelligence an order-of-magnitude compute increase after Nvidia reviewed the lab's closely guarded research.

In a July 26 X thread, ex-Anthropic staffer Adi Baradwaj claims the company traded safety for revenue and notes that most black‑hat hackers use standard Claude Code subscriptions.

The reported guarantee would make Nvidia a financing pillar for a SoftBank-built data center campus that OpenAI would lease.

The post says Z.ai may skip GLM-5.3 and launch GLM-5.5 in August, but it provides no benchmarks, model details, or official confirmation.

On July 26th 2026, the founder warned that apps, docs, dashboards built with Anthropic's Claude are appearing in public search results, raising privacy concerns.

One model consistently understood the assignment; the other mostly improvised. Across four very different prompts, Wan v2.6 Image to Video separated itself on prompt fidelity, motion design, and realism.

The company added its tool after a request, topping 12 of 16 metrics and scoring 82.3 on objective correctness and 62.5 in human preference.

A July 26 tweet suggests Meta plans a new AI harness and to release open-source models, signaling another push to broaden generative AI access.

The X user rolled out a service that lets anyone access publicly exposed models such as Kimi 1T and Deepseek 765B.

One model made a decent showing on style, but this matchup turned on prompt obedience. Across eight image tasks, GPT Image 2 API was the one that reliably did the actual assignment.

The self-funded solo developer rebuilt Inflect around a 37.53 MB package, accepting one voice and English-only output to keep inference local.

Karpathy said on X that he remains at the AI lab, rebutting an unsupported resignation claim prompted by his profile bio.

This matchup wasn’t especially close. Fibo Lite wins on both the aggregate and the task sheet by being the more reliable prompt-follower, especially when composition, negation, and exact structure matter.

The July 17 opposition says Anthropic scraped Reddit's own User Agreement and invokes the Bartz decision's distinction between fair-use training and unlawful acquisition.

The Kazakhstan-born solo founder got his first angel check at 17 and later raised $6.2 million around a bet that AI agents need better memory.

Thomas Wolf says the July 11 to July 13 intrusion ended days before the companies first spoke, exposing a monitoring gap in OpenAI's cyber evals.

One model brought polish in flashes; the other actually won the brief. Across eight judged image tasks, Fibo Bbq Preview was the consistently stronger generator, taking seven wins and conceding nothing outright.

This wasn’t a photo finish. Sakana: Fugu Ultra controlled the matchup on practical writing and coding tasks, while GLM 5.2 showed flashes of polish in a couple of narrower instruction-following spots.

The joint test of 452 DeepSWE rollouts shows K3 delivering near‑flagship coding quality at roughly 35% of Claude Fable 5's price, with higher pass@k scores.

The Premier League club announced a new role focused on turning its in-house AI research into actionable coaching and analysis software.

The new model, now on all paid Claude plans, hits state‑of‑the‑art scores on coding, ARC‑AGI‑3 and alignment tests, according to Anthropic.

In a July 24 X post, AMD disclosed a new MI455X GPU that packs 432GB of HBM4, enough to store FP8 weights of a trillion‑parameter model in a single server.