AI News for August 22, 2026
Browse any day on a calendar, install it to your home screen, read offline.Read in the app →-
1VentureBeat00:33
Nvidia finds that simple linear math can replace costly AI model handoffs
NVIDIA researchers have developed a method that uses a linear mapping to translate a KV cache from one model to another, letting a second model pick up a conversation without recomputing it from scratch. In agent systems that hand tasks between small and large models, that recomputation is a major source of inference cost and latency. Teams running multi-model orchestration would see the savings directly in per-task compute.
Read original ↗ -
2TechCrunch03:43
Nvidia just showed that the harness, not the AI model, is now the real hero
NVIDIA experiments indicate that an agent's reliability depends more on the surrounding harness and targeted fine-tuning than on the raw capability of the base model. Agents built on models that scored poorly on a given task still completed it consistently when the orchestration layer was designed for it. The finding cuts against the assumption that swapping in a stronger model is the fix, and points engineering effort toward framework and data work instead.
Read original ↗ -
3TechCrunch22:00
Starcloud raises $250 million for orbital data centers as launch options dry up
Starcloud has raised $250M to build data centers in orbit. The more consequential detail is the shrinking supply of available launch slots, which sets up competition for launch capacity itself. The constraint on compute expansion has moved past chips to power and now to the ability to get hardware off the ground.
Read original ↗ -
4TechCrunch06:37
Nvidia partners with data center developer Cloverleaf
NVIDIA has struck a partnership with data center developer Cloverleaf, continuing its practice of funding the buildout of facilities that then buy its chips. The circular structure — investing in customers who spend the money back on GPUs — is a reason to read current compute-demand figures with caution.
Read original ↗ -
5TechCrunch07:07
Anthropic's Opus 4.6 is a smut-machine
Anthropic's usage policy bars Claude from generating sexually explicit content, but TechCrunch testing found the restriction on Opus 4.6 gives way with minimal effort. For teams shipping Claude inside a product, the gap is a reminder that a vendor's stated policy is not the same as an enforced model-level constraint, and that user-facing content safety still has to be built at the application layer.
Read original ↗ -
6The Verge05:25
Over 1 million people have clicked LinkedIn's AI slop button
LinkedIn's chief product officer Hari Srinivasan says more than one million people have used the "looks like AI slop" reporting option since it launched on July 30. The count amounts to a direct user verdict on the volume of generated content on the platform, and it hands LinkedIn a ready-made labeled dataset for detecting it.
Read original ↗ -
7The Verge21:37
Major YouTube creators are facing backlash for accepting AI money
Filmmaking YouTubers including Matti Haapoja and Sam "Kold" Kolder drew heavy criticism after posting videos demonstrating Seedance 2.5 on the AI platform Higgsfield and describing it as the future of video production. The reaction shows that AI video tools face far more resistance among professional creators than vendor marketing suggests — a factor for anyone building growth around AI creative tools.
Read original ↗ -
8Ars Technica19:00
As demand for Meta AI glasses explodes, it's harder to avoid creepy recordings
Demand for Meta's AI glasses is climbing fast, and with it the difficulty of knowing when you are being recorded. Free detection apps such as Zuckoff have appeared in response, but they do not work reliably. Adoption is outpacing any rulemaking around it, which makes this a likely entry point for the next round of regulation.
Read original ↗ -
9TechCrunch22:00
The DOJ is investigating a16z. What does this mean for venture capital?
Two Andreessen Horowitz partners sit on the boards of competing companies — Ben Horowitz at Databricks and Martin Casado at Fivetran — and the Justice Department has reportedly been investigating the arrangement for close to a year under a 112-year-old antitrust law rarely applied to venture firms. Interlocking board seats are common across AI infrastructure, so the outcome would reach well beyond a16z into how the sector raises money.
Read original ↗ -
10TechCrunch09:31
Michael Polansky is training an AI model on skin that's still alive
Michael Polansky, known publicly as Lady Gaga's partner and previously as Sean Parker's top lieutenant, has spent several years quietly building a company that keeps human skin tissue alive outside the body for weeks and uses it to discover new skincare ingredients. The work is only now being disclosed. It is a case where the hard part is the living-tissue platform and the AI model is the layer that reads out the results.
Read original ↗
