Kagi News - AI

AI power startup Fermi loses CEO and CFO

Mon, 20 Apr 2026 15:35:42 +0000

Fermi, a startup pitching an AI-focused data center and power campus in Texas, has lost its chief executive and chief financial officer, according to reports from the Financial Times, TechCrunch, and Fortune. The leadership shake-up comes during a difficult stretch for the company, which former U.S. Energy Secretary Rick Perry co-founded and has been trying to move its Texas project forward. The reports describe a company struggling to advance its Texas AI campus plans while also losing a planned $150 million investment from Amazon, adding pressure to an already capital-intensive business model. For AI readers, the episode is a reminder that excitement around data centers and power still depends on execution, financing, and leadership—not just demand growth.

Highlights:

Amazon setback: Fermi lost a planned $150 million investment from Amazon, removing a major source of funding for its build-out plans.
Texas project: TechCrunch reported that Fermi had been running into headwinds around its proposed AI campus in Texas, underscoring how hard it is to combine power development with large-scale AI computing sites.
Political link: Former U.S. Energy Secretary Rick Perry co-founded Fermi, giving the startup a higher public profile than many early-stage infrastructure ventures.
Reset framing: Fortune, via Google News, described the management shake-up as part of a "2.0" reset.

Perspectives:

Financial Times: The outlet portrayed Fermi as a data-center hopeful now hit by management departures and the loss of Amazon’s $150 million investment. (Financial Times)
TechCrunch: TechCrunch highlighted the sudden departure of the CEO and CFO and tied it to operational headwinds around the company’s Texas AI-and-power campus. (TechCrunch)
Fortune: Fortune described the company as financially strained and cast the leadership changes as part of a "2.0" reset effort. (Fortune)

Sources:

Shares in data centre hopeful Fermi plunge as top executives quit - ft.com
CEO and CFO suddenly depart AI nuclear power upstart Fermi - techcrunch.com
Financially struggling, AI power startup Fermi loses its CEO and CFO for ‘2.0’ reset - Fortune - google.com

AI researchers unveil safer, leaner, more reliable LLM methods

Mon, 20 Apr 2026 14:16:40 +0000

A broad wave of AI research published on April 20-21 focused on making large language and multimodal models more reliable, efficient, and useful in real deployments, not just stronger on static benchmarks. Across dozens of papers, researchers proposed ways to reduce hallucinations, improve safety in constrained or multi-turn settings, strengthen long-context memory, and cut the token, memory, and latency costs of serving models at scale. The mood of this research cycle is notably practical. Several teams released benchmarks designed to expose failure modes that ordinary leaderboards can miss, including multimodal misinformation, sarcasm, geometry representation sensitivity, structured safety bypasses, long-horizon personalization, confidence validity, and mental-health interaction risks. At the same time, many papers pointed to real progress, showing that better decoding, retrieval, memory design, reinforcement learning, and domain-specific adaptation can improve accuracy or robustness without always requiring bigger models. That is an encouraging sign for making AI systems more dependable and more broadly accessible.

Highlights:

Safety gaps: Researchers found that forcing harmful prompts into multiple-choice formats can sharply increase policy-violating answers compared with open-ended prompts, suggesting current safety evaluations miss structured-decision risks.
Medical promise: A Nature paper described a multi-agent framework that combines large language models with medical flowcharts for self-triage.
Efficiency push: Infrastructure work included a cross-datacenter KV-cache architecture for serving LLMs at scale.
Multilingual gains: Several papers addressed language inequities by improving multilingual prompting and low-resource language adaptation.

Perspectives:

Safety researchers: They argue that common open-ended safety tests understate risk because refusal behavior often collapses when the same harmful request is recast as a forced-choice multiple-choice question. (arXiv)
Memory-system researchers: They argue that deployed LLM memory should be treated as a lifecycle management problem, with promotion, retention, and eviction decisions, rather than as a simple write-and-retrieve store. (arXiv)
Localization researchers: They argue that AI safety guardrails need language- and region-specific tuning, reporting that a Taiwan-focused model substantially improved practical guardrail performance over a generic base model. (arXiv)
Long-context evaluation researchers: They report that current frontier models often fail to track evolving user preferences across six-month conversation histories, with belief-update failures remaining the main bottleneck. (arXiv)
Industry analysts: GlobeNewswire market items framed custom LLM training platforms and data-lineage tools as growing commercial categories, underscoring enterprise demand for governance and model-building infrastructure. (Google News / GlobeNewswire)
Interpretability researchers: They suggest that many important properties of reasoning and steerability are easier to read from internal activations and geometric stability than from model outputs alone. (arXiv)

Sources:

Amazon expands Anthropic deal with up to $25 billion investment commitment

Mon, 20 Apr 2026 22:31:00 +0000

Amazon and Anthropic said they are expanding their AI partnership through a package that combines up to $25 billion in new Amazon investment with a long-term infrastructure commitment under which Anthropic plans to spend more than $100 billion on Amazon Web Services technologies. The companies also said the collaboration will give Anthropic up to 5 gigawatts of new compute capacity, with Anthropic using Amazon’s Trainium chips to build and run future Claude models. The deal shows how the AI race is increasingly being shaped by access to chips, power, and cloud capacity, not just software. Several reports said Anthropic pursued the larger arrangement after outages and rising demand exposed the need for more reliable computing resources, while Amazon is strengthening its position against Microsoft-backed OpenAI and Google-backed rivals by tying a leading model developer more closely to AWS. The agreement gives Anthropic a clearer path to scale Claude and gives Amazon a very large cloud customer as AI infrastructure becomes one of the industry’s main battlegrounds.

Highlights:

Ten-year term: Reuters reported that the arrangement runs for 10 years, adding duration and visibility to the cloud commitment beyond the headline investment figures.
Hardware focus: Anthropic said the expanded collaboration centers on Amazon’s custom Trainium infrastructure, while Amazon said the pact supports both model training and deployment on AWS.
Stock reaction: Investor-focused coverage said Amazon shares rose after the announcement, reflecting market interest in the commercial upside of locking in AI infrastructure demand.
Circular economics: TechCrunch described the arrangement as a circular AI deal because Amazon is investing capital while Anthropic simultaneously commits massive spending back to AWS.

Perspectives:

Anthropic: The company framed the agreement as a way to secure up to 5 gigawatts of compute and expand access to the infrastructure needed for future Claude models. (Anthropic)
Amazon: Amazon presented the pact as a deeper strategic collaboration that strengthens AWS as a platform for developing and serving frontier AI models. (About Amazon)
Financial Times: The FT described the deal as a response to compute shortages and outages that pushed Anthropic to secure more chips and power for its models. (Financial Times)
Axios: Axios cast the announcement as part of the wider compute wars, arguing that access to infrastructure has become a competitive weapon among leading AI companies. (Axios)

Sources:

Anthropic, Amazon Tighten Bond in $5 Billion Investment and Computing Deal - WSJ - google.com
Amazon Plans to Invest Up to $25 Billion in Anthropic - nytimes.com
Anthropic and Amazon agree $100bn AI infrastructure deal - ft.com
Anthropic takes $5B from Amazon and pledges $100B in cloud spending in return - techcrunch.com

Moonshot AI launches Kimi K2.6 open-weight model

Tue, 21 Apr 2026 00:20:06 +0000

Moonshot AI has released Kimi K2.6, a new open-weight AI model family available through Kimi Chat, with published weights and API access also described in specialist coverage. Across specialist outlets, K2.6 is described as a 1 trillion-parameter mixture-of-experts model with 32 billion active parameters, native multimodality, and attention optimizations. Moonshot is positioning it against Anthropic’s Claude Opus 4.6 and OpenAI’s GPT-5.4. One bright spot for developers is how quickly the model seems to have reached the broader tooling ecosystem. One newsletter report said it had day-one support in platforms including vLLM, OpenRouter, and Cloudflare Workers AI, while Reddit discussion quickly turned to local deployment, quantization, and whether it could replace premium closed models. The release adds another open-model contender in agentic coding and long-horizon tasks, an area where newsletter and analyst coverage says Chinese open and semi-open labs are gaining momentum and giving developers more options beyond closed frontier systems.

Highlights:

Benchmark claims: Moonshot said K2.6 scored 54.0 on Humanity's Last Exam with tools, 58.6 on SWE-Bench Pro, 76.7 on SWE-bench Multilingual, 83.2 on BrowseComp, 50.0 on Toolathlon, 86.7 on CharXiv with Python, and 93.2 on Math Vision with Python.
Agent scale: One newsletter summary said the model was built for long-horizon execution with more than 4,000 tool calls, runs lasting over 12 hours, and as many as 300 parallel sub-agents.
Model variants: K2.6 launched in four versions: Instant for speed, Thinking for deeper reasoning, Agent for research and document tasks, and Agent Swarm for large-scale search, batch work, and long-form output.
Local hardware: Community posts showed both excitement and practical limits, including a GGUF Q4_X quantization that reportedly needs more than about 584GB of combined RAM and VRAM, alongside separate discussion of what hardware would be needed to run the model locally at 25-30 tokens per second.

Perspectives:

Moonshot AI: Moonshot presented K2.6 as open-source state of the art for coding and agentic workloads, emphasizing benchmark leadership and broad availability through chat, APIs, and published weights. (TestingCatalog)
AI newsletter analysts: Latent Space described K2.6 as a refresh that helps Moonshot keep pace with Claude Opus 4.6 and maintain its lead among Chinese open-model labs in 2026. (Latent Space)
Developer community: One Reddit user said Kimi K2.6 was the first model they would confidently recommend as an Opus replacement for customers, estimating it could handle about 85% of those tasks while offering vision and strong browser use. (Reddit)
Industry observers: Smol.ai framed the release as evidence that Chinese open and semi-open labs are building competitive momentum in coding and agent models, especially through open weights and broad platform support. (smol.ai)

Sources:

Moonshot AI’s Kimi K2.6 launch challenges Anthropic’s AI dominance - Crypto Briefing - google.com
Moonshot AI launches Kimi K2.6 on Kimi Chat and APIs - testingcatalog.com
Kimi K2.6 is a legit Opus 4.7 replacement - reddit.com
[AINews] Moonshot Kimi K2.6: the world's leading Open Model refreshes to catch up to Opus 4.6 (ahead of DeepSeek v4?) - latent.space
not much happened today - smol.ai