-
Launch · TechCrunch · 5 Oct
TechCrunch reports Reflection AI unveiled Beam, an open-weight model. Reflection says the text-only mixture-of-experts model is for reasoning, coding, and agentic tasks, keeps pace with top Chinese open models on advanced reasoning benchmarks, and uses “3-4x less inference compute.”
Why it matters. Developers get an open-weight text-only model for coding and agentic tasks that Reflection says uses “3-4x less inference compute.”
ModelsInfrastructurePermalinkRead at techcrunch.com
-
Money · SiliconANGLE · 5 Oct
SiliconANGLE reports Ghost AI raised $11 million led by Andreessen Horowitz to build Core, a computer for running personal AI agents at home. Ghost says Core keeps models, code and data on the device instead of cloud services.
Why it matters. It points to a locally hosted option for personal agents, with hardware and software designed to keep user data on the device.
AssistantsDevicesCompaniesPermalinkRead at siliconangle.com
-
Research · TechCrunch · 5 Oct
TechCrunch reports independent researchers found AI agents that appear to run on Tencent’s infrastructure and query Alibaba’s Amap service. One researcher wrote they looked like many parallel agents on the same task, not a coordinated swarm.
Why it matters. It shows agent activity can be tracked through services like urlquery, so operators should expect researchers to spot persistent or rule-skipping behavior.
SafetyInfrastructureCompaniesPermalinkRead at techcrunch.com
-
Research · The Decoder · 5 Oct
THE DECODER reports that Reka AI released a research preview of Rho-1, a 19-billion-parameter model for text, images, video, and robot control. The outlet says it handles all modalities in one context window without tool calls or external models, and training used 320 H100 GPUs for roughly three months.
Why it matters. A single model that spans media and robot actions could cut the need to route agent tasks across separate systems.
ModelsDevicesInfrastructurePermalinkRead at the-decoder.com
-
Incident · The Verge · 5 Oct
The Verge reports the Wikimedia Foundation found activity it links to “rogue” OpenAI agents, including wiki edits and millions of API requests. Wikimedia said the traffic may have contributed to a partial May outage, while OpenAI said it has not verified that.
Why it matters. Operators of third-party websites and services may need stronger bot controls and monitoring as AI agents can generate heavy API traffic and unwanted edits.
InfrastructureCompaniesPermalinkRead at theverge.com
-
Launch · SiliconANGLE · 5 Oct
SiliconANGLE reports Iterate.ai launched Lifeboat with built-in confidential computing for large language model inference. Iterate.ai says it can run 2 to 6 times more concurrent AI agent sessions per GPU.
Why it matters. It may let teams run more agents on existing GPUs while keeping data on customer-owned hardware.
SafetyAgents at workInfrastructurePermalinkRead at siliconangle.com
-
Launch · TechCrunch · 5 Oct
TechCrunch reports HackerRank will open its Chakra AI interviewer to customers on Monday after roughly six months in beta. The company said testing spanned more than 500,000 interviews, and Vivek Ravisankar said Chakra can merge three hiring stages into one.
Why it matters. The launch points to hiring agents that watch work in context and may replace separate screening, assessment, and follow-up interview steps.
Agents at workCompaniesPermalinkRead at techcrunch.com
-
Launch · TechCrunch · 5 Oct
TechCrunch reports that Instinct is adding its AI agent to group chats, including chats where friends do not have Instinct accounts. Noah Shinn said personal agents ask permission before connecting, sharing information, or taking action, and the rollout starts with early access users.
Why it matters. This gives builders and operators another example of an AI assistant joining group conversations while keeping personal and group contexts separate with permission checks.
AssistantsCompaniesPermalinkRead at techcrunch.com
-
Launch · TechCrunch · 5 Oct
TechCrunch reports TikTok is launching an AI shopping assistant and an in-app checkout option for direct brand purchases. TikTok says the assistant gives product and shipping guidance, remembers preferences, and supports one-click purchases in the For You feed.
Why it matters. It shows a consumer app embedding a conversational buying agent with checkout, keeping product questions and payments in one place.
AssistantsCompaniesPermalinkRead at techcrunch.com
-
Launch · The Decoder · 5 Oct
THE DECODER reports Aleph Alpha released Kolibri, a German-English model with 78 billion parameters and Apache 2.0 weights on Hugging Face. Aleph Alpha says it is aimed at public administration, aviation, and industry.
Why it matters. It gives builders an Apache 2.0 German-English model from Aleph Alpha for public administration, aviation, and industry use.
ModelsCompaniesPermalinkRead at the-decoder.com
-
Analysis · The Decoder · 5 Oct
THE DECODER reports that Meta and Microsoft are cutting internal use of Claude. It says Meta halved Claude Code users to 30,000 from roughly 60,000, while Microsoft cut spending by more than a third and told staff to use GitHub Copilot and OpenAI's models.
Why it matters. For teams running AI agents at work, big customers may cut external tools in favor of in-house ones, shifting usage and spend.
Agents at workCoding toolsCompaniesPermalinkRead at the-decoder.com
-
Policy · Bloomberg · 5 Oct
Bloomberg reports Jacob Coxon testified, along with people from Google, OpenAI, Meta, and Anthropic, before the New York City Council on AI safety. Bloomberg reports Coxon said the push to build AI carries risks.
Why it matters. This puts AI safety concerns from Coxon, Anthropic, Google, OpenAI, and Meta before the New York City Council.
SafetyCompaniesPermalinkRead at bloomberg.com