Large Language Models

Today's Stories

1. 1

   Honeywell maps six workflows for AI-driven autonomous asset optimization

   Honeywell is using AI-driven workflows and agent-based automation to help manufacturers address labor shortages, skilled-worker scarcity, and product complexity. At the 2026 Honeywell User Group Americas Conference in Phoenix last month, Omar Sayeed, digital reliability leader at Honeywell, outlined six core workflows—asset surveillance, root-cause analysis, prescriptive recommendations, dynamic risk-based maintenance strategy, production-reliability trade-off evaluation, and field-worker support—that enable autonomous asset optimization. Manufacturers face tightening labor markets, workforce problems, supply chain issues, and pressure to meet deadlines. Connected, AI-enabled maintenance workflows help companies automate and reduce human intervention while maintaining asset reliability and performance. Honeywell's recent acquisitions of turbo machinery manufacturer Sundyne and Compressor Controls Corp. underscore the company's strategic focus on expanding asset reliability services and positioning AI agents as a lever for manufacturers to operate at scale.

   Success requires a foundation of robust data collection, strong control infrastructure, and analytics platforms. Sayeed emphasized that the biggest barrier for many facilities is fragmented or poor-quality data—AI cannot help without it. Organizations must establish centralized, standardized work processes and move progressively from predicting failures toward prescribing actions, ultimately enabling autonomous control decisions with minimal human involvement.

2. 2

   Google delays Gemini launch after missing internal benchmarks

   Google has postponed the launch of its Gemini AI technology because the system failed to meet internal performance targets set by the company. The delay signals that even leading AI labs face technical hurdles in scaling up their largest models. For businesses planning to adopt advanced AI, it suggests real-world deployment timelines may slip beyond public expectations.

   No specific launch date has been announced yet; the company is working to close the performance gap before making the technology available.

3. 3

   Google's Gemini 3.5 Pro delayed months, coding lags expectations

   Gemini 3.5 Pro, Google's flagship AI model, is running months behind schedule because it fell short of Google's internal expectations for coding capabilities. The delay signals potential weakness in Google's AI development pipeline at a time when the company faces intense competition in generative AI. For businesses and developers relying on Google's AI tools, this suggests a timeline slip for accessing an expected next-generation capability.

   The article does not specify when Gemini 3.5 Pro is expected to launch or the exact duration of the delay beyond "months."

4. 4

   Google Stock Slides as Gemini 3.5 Pro Release Delayed

   Google's stock price fell following a report that the company is behind schedule in delivering Gemini 3.5 Pro, its most powerful flagship AI model. The delay comes as AI rivals are making progress with their own models, creating competitive pressure in the high-stakes market for advanced AI capabilities. For businesses evaluating AI partners, timing and capability gaps between vendors matter.

   The article does not specify when Gemini 3.5 Pro is now expected to ship or by when the delay will be resolved.

5. 5

   Agents Need Better Web Search—Full Documents, Not Snippets

   A analysis of how AI agents search the web shows that traditional search APIs returning only snippets and links force agents to fetch and clean raw HTML on every query. Running the same question through three retrieval methods—plain search results, neural search, and a pre-indexed web service—revealed that looping through snippets costs roughly 4× more tokens than accessing full pre-processed documents. For a research task requiring multiple searches, agents repeatedly repay a "retrieval tax" on content they've already seen. As coding agents and LLM-powered systems move to production, validation shifts from offline tests to live telemetry—meaning agents must handle real-world queries they were never trained on. Web search is how agents stay current with markets, news, and organizational changes, but today's search APIs waste tokens on overhead instead of reasoning. For teams building agent loops (research pipelines, GTM briefings, market monitoring), the choice of search tool directly controls cost and speed—not the model or prompt.

   Seltz, a purpose-built web index for agents, returns complete structured documents (people profiles, full article text, Wikipedia entries) in a single call rather than forcing agents to reconstruct content from snippets. It works best for deep-dive lookups after discovery is done; chaining discovery queries (open web search) with deep-dive queries (indexed full documents) lets each loop iteration use the right retrieval method for what it's actually trying to do.

6. 6

   Kimi K3 open model matches GPT-5.6 Sol, Claude Fable 5 in benchmarks; China AI prices rise

   Kimi launched K3, a 2.8 trillion parameter open model with a one million token context window. In Kimi's benchmarks, K3 trails only Claude Fable 5 and GPT-5.6 Sol but beats all other tested systems including Claude Opus and Chinese rival GLM-5.2. Independent testing by Artificial Analysis scores K3 at 57 on the Artificial Analysis Intelligence Index, placing it fourth behind Fable 5 (60), GPT-5.6 Sol (59), and Opus 4.8 (56). Full model weights are scheduled for release by July 27. K3 signals the end of ultra-cheap Chinese AI. Kimi's pricing—$0.30 per million input tokens (with cache hit) and $15.00 for output—is nearly 19 times higher than its predecessor K2.6 ($0.16 input, $4.00 output). Chinese providers overall are raising prices for frontier models. At $0.94 per task, K3 lands in the same range as GPT-5.6 Sol ($1.04) but costs roughly half of Claude Opus 4.8 ($1.80), positioning it as a competitive midrange option rather than a price-leader.

   K3 shows higher hallucination (51 percent) than its accuracy improvement (46 percent, up from 33 percent on the AA-Omniscience Index), and it uses a mixture-of-experts architecture activating only 16 of 896 experts. The model is available now via Kimi.com, mobile apps, and Kimi Code; a planned Kimi Hosted Agent platform with isolated environments for long-running tasks is in waitlist signup.