SK Telecom's A.X K2 landed July 29 at 688B parameters, and LG's K-EXAONE 2.0 followed July 31 at 750B. Washington answered by banning new authorizations for foreign-made humanoids and robot dogs.
SK Telecom released A.X K2 on July 29, a 688B-parameter open-weight MoE with 33B active parameters trained from scratch in FP8. LG AI Research followed on July 31 with K-EXAONE 2.0 at 750B total and 37B active, roughly tripling its prior 236B model.
Google DeepMind detailed Gemini Robotics ER 2 on July 30 with step-change gains in video understanding and tool orchestration. The shift from single arm to multi-robot coordination is the part builders should read twice.
Iran warned Saturday it would strike energy fields in other countries if the United States launches fresh attacks, a threat aimed at Gulf oil and gas infrastructure. Embassies issued citizen advisories the same day.
Koda publishes a fact-checked AI intelligence briefing every morning: site, email, podcast, and video. Founding sponsors lock today's rates for 6 months as the audience compounds.
View the media kit →SK Telecom released A.X K2 on July 29 as a 688B-parameter open-weight MoE model with 33B active parameters, trained from scratch in FP8 using the company's Sparse Gated Attention architecture. LG AI Research followed on July 31 with K-EXAONE 2.0 at 750B total and 37B active parameters, roughly tripling the size of...
Continue reading arrow_outwardSK Telecom's 688B-parameter A.X K2 on July 29 and LG AI Research's 750B K-EXAONE 2.0 on July 31 put Korea in the open-weight frontier tier within a single week.
Iran threatened Saturday to strike energy fields in other countries if the US attacks again, and US embassies in roughly 10 Middle East countries told Americans to consider departing.
Mood reads Fear, with Gulf energy infrastructure under explicit threat and Washington closing the FCC authorization door on a category of imported hardware.
Robotics is short two things at once: the US just restricted foreign-built humanoids while startups pay private chefs to wear brim cameras in strangers' kitchens because training data does not exist yet.
A.X K2 was trained from scratch in FP8 using SK Telecom's Sparse Gated Attention architecture with 33B active parameters; K-EXAONE 2.0 runs 37B active out of 750B total, roughly tripling LG's earlier 236B model. Two national-champion labs releasing at this scale days apart signals coordinated ambition rather than coincidence. The open-weight frontier is no longer a two-country conversation.
The US barred new authorizations for foreign-made humanoid robots, quadruped robot dogs and solar inverters, citing network-connected hardware built abroad; affected devices cannot legally be sold as new. That protects a domestic robotics market that still lacks the egocentric training data its policies require, which is why Shift is paying to film private chefs in San Francisco homes. Restricting supply does not solve the data bottleneck.
OpenAI's Building abundant intelligence essay argues commodity-priced intelligence demands industrial-scale compute, arriving days after the company cut GPT-5.6 Luna prices 80 percent. Meanwhile Google DeepMind's Gemini Robotics ER 2, detailed July 30, moves embodied reasoning from a single arm to multi-robot collaboration and tool orchestration. Both bets assume the physical and compute buildouts keep pace with the models.
Deep Dive
7.1/ 10
An Apple Design Award winner that solves exactly one problem well: naming the things in front of you in a language you barely speak.
CapWords is for travelers, beginners and visual learners who want concrete-noun vocabulary tied to real places, not a curriculum. It is genuinely pleasant and free to try, and the photo-as-memory-anchor trick works better than generic flashcard decks for objects, signs and menus. If you need grammar, structure or serious speaking reps, this is a companion app, not your main one.
“Fun, engaging and very easy to use, especially good for beginners and visual learners, but the limited free scans are frustrating.”SpeakTwice app review
This free, read-only MCP server exposes Google Ads, Meta Ads and LinkedIn Ads data directly to your assistant, so you can ask about spend and performance without exporting CSVs. Read-only access is the point: the agent can analyze but cannot change bids or budgets by accident. Best first test is a weekly cross-channel spend question you currently answer by hand.
Try it arrow_outwardInstead of one-off prompts, Claude Tag sits in a channel with shared context so follow-ups, coding tasks, analysis and incident support all build on the same thread history, with enterprise governance controls attached. Try it in a single low-stakes channel first, such as an on-call or triage room, and see whether the retained context actually reduces re-explaining. If it does not, the failure mode is obvious within a week.
Try it arrow_outwardManus now ships a macOS and Windows desktop app that runs its agent on your local machine rather than in a remote sandbox, which matters when the work involves local files, installed apps and credentials you do not want uploaded. Scope it to a throwaway user account or project folder before you point it at anything real. Local execution shifts the risk from data transfer to what the agent can touch on disk.
Try it arrow_outwardIt surfaced this week on There's An AI For That, alongside the current crop of prompt-to-app builders, and has not yet been picked over by the usual review cycle. Spend 20 minutes on it with a project you already know the shape of, so you can judge output quality against a baseline instead of being impressed by speed. Check the export path before you invest: what you cannot get out, you do not own.
Try it arrow_outwardBenchmark the Korean models against your incumbent. Run your top ten production prompts through A.X K2 and K-EXAONE 2.0 and compare cost per task at 33B and 37B active parameters rather than trusting headline totals.
Spend 20 minutes on Gemini Robotics ER 2. Read the multi-robot coordination section with a project in mind, then sketch where fleet-level tool orchestration would replace a single-agent loop in your own system.
Wire an agent into where work already happens. Test Claude Tag as a resident Slack agent, Connector.wtf for ad account access, and Manus My Computer for desktop execution, then keep only the one that removes a manual step.
Two sovereign labs shipped frontier-scale open models 48 hours apart, and the policy response was aimed at physical robots rather than model weights. That asymmetry is the story: software capability now diffuses faster than any authorization regime can gate it, while robotics stays bottlenecked on hardware supply and real-world data that startups are literally paying to film in kitchens. Builders should assume capable open weights keep arriving from more places, and that the scarce inputs are compute commitments and proprietary task data. Test the new models this week, and put your moat where the ban lists cannot reach.
AI analysis, world news, markets, and tools. One briefing, delivered free.