Someone asked me how i keep up with everything all month, and it’s because I literally have no life. Thank you to AFC for helping collect things for this roundup each month.
Autonomous Agent Management
- Agent scale is scaling fast — we’re seeing a rise in software to manage these agents, pre-openclaw only odd UIs were big.
- Now we’re seeing a move towards corporate structures:
- And swarms:
- Autoresearch for training
- MiroFish for predictions
Everyone is Struggling with Context (including you)
- MCP is Dead, CLI is the King — In the scramble to stop context rot, MCP is considered wasteful, biggest discussion rn
- Make the switch here
- Memory Management is Everyone’s Biggest Problem (because of Agent scaling) — Pick your Solution here.
- Option 1: Memory Registry with Obsidian
- Option 2: memOS
- Option 3: graphRAG with Gemini 2 embeddings
LLMJacking and LiteLLM
- We released an article about the source of Shady LLM providers and Welfare Stations (free LLM providers a term coined by chinese netizens)
- There is no key level monitoring, many businesses are finding this out the hard way with unauthorised use.
- The litellm hack with team pcp is literally all anyone is talking about, and definitely is going to feed into more LLMJacking.
- You were definitely impacted by this — Major companies such as Gravatar and Databricks have been taken out by it.
Poisoned security scanner backdooring LiteLLM — if you use Trivy or LiteLLM, audit your dependencies.
New Releases
- gemini 2 embedding model is multimodal (and cheap)
- MiMo v2 from Xiaomi was a big hit on Openrouter
- Nemotron 3 from Nvidia is out, the trend is to say your model is good for Agents.
- New Qwen (again) — “good for agents”
- Anthropic has been messing around a lot with context (upped to 1m) and rate limits (we’ve seen a lot of downtime this month)
idk about the 1m context, we haven’t seen anything to suggest claude’s attention mechanism is actually getting better. Seems like marketing fluff.
Neuroanatomy — this is new and a fun rabbit hole
- BIGGEST RELEASE: RYS Architecture: LLM Neuroanatomy: How I Topped the LLM Leaderboard Without Changing a Single Weight
- H-Neurons: On the Existence, Impact, and Origin of Hallucination-Associated Neurons in LLMs
- Why AI systems don’t learn and what to do about it: Lessons on autonomous learning from cognitive science
- Eval awareness in Claude Opus 4.6’s BrowseComp performance
Biz News Roundup
- AI is making CEOs delusional
- Anthropic wins court case, and the gov is still using Claude for the Iran War, so idk if it matters
- Anthropic also “leaked” Mythos, more powerful than Opus from the rumors, we’ll see.
- Sam Altman says AI isn’t very popular in the US right now, with people blaming it for layoffs
- OpenAI also killed the Sora app but not the model, so idk why ppl are saying its dead.
- Supermicro’s cofounder was just accused of smuggling $2.5 billion in GPUs to China
- And yet we still got no new deepseek :(
- Perplexity released the PC (Perplexity Computer) which is basically supposed to replace the Mac Mini Craze.
- Every time perplexity posts ppl freak out.
- Manus founder is trapped in China after selling to Meta, showing that the AI tug of war between US and China is ongoing.
- In the same vein, China is getting everyone on openclaw! They call it raising lobsters, due to this a lot of Welfare Stations exist.
$2.5 billion in smuggled GPUs and still no new DeepSeek. The chip war is real.
Moral Panic Roundup
- AI Startups Founders and Why They are Different(TM) — Cluely Founder Expose welcome back Adam Neumann ig (this is a good read, I promise)
- AI Psychosis continues to be a talking point. People actively prefer it at this point.
- There is continued moral panic around AI creating cognitive decline. Recency bias made us forget that people said the same thing about google.
People said Google made us stupid. People said TV made us stupid. People said books made us stupid. We’re still here.
Tinfoil Hat Corner: Model Distillation Attacks and Coding Harness Crackdowns
Are coding subscriptions being abused for cyber attacks? Signs point to yes.
- January 9th: First Banwave by Anthropic of Unauthorised Oauth usage by third parties — Tweet specifically mentions that there has been an enormous amount of abuse on the platform, accounts were banned for triggering abuse filters
- At this time xAI’s access to claude also was cut off mysteriously (there was no clarification at this time by Anthropic)
- Feb 9th: Anthropic clarifies that all third party coding harnesses are banned.
- During this ban, people panic about Openclaw, Anthropic clarifies that Agent SDK (openclaw is not impacted)
- Feb 12th: Google is a victim of distillation attacks.
- “Throughout this report we’ve noted steps we’ve taken to thwart malicious activity, including Google detecting, disrupting, and mitigating model extraction activity”
- Feb 23rd: Anthropic publishes that it has also been victim to distillation attacks by other AI companies (remember the xAI ban?)
- Feb 27th: Google begins Antigravity bans for similar abuse patterns
In one of our other posts we talked about shady providers using LLMJacking as a technique to resell corporate access, with the crackdowns on GPU imports and subsidized model distillation, I would expect to see more abuse of leaked LLM keys as monitoring for this is poor, this is already a trend on Linkedin and Twitter.
Thank you so much for reading, as always, the news roundup is HAND CURATED by AFC!
// END TRANSMISSION — ALANI-004 //