Tech / AI / IT Monitor

Executive Summary

The last 24 hours were dominated by OpenAI's launch of GPT-5.6 (Sol and Terra models) under forced US government limited preview, sparking widespread debate about regulatory capture and AI access restrictions. Anthropic faced intense community backlash over anti-competitive accusations and the apparent pullback of its "Fable" model. Meanwhile, the open-source/local AI ecosystem countered aggressively: Nous Research released Mixture of Agents 2.0 in Hermes Agent, the Ornith 35B agentic coding model demonstrated self-verification on a single DGX Spark at near-lossless FP8, and DeepSeek open-sourced DeepSpec, a speculative decoding training framework. The overarching tension is between gated frontier access and an increasingly capable open/local alternative stack.

Key Events

• OpenAI launches GPT-5.6 Sol & Terra under US government limited preview — Sam Altman announced Sol (frontier-tier) and Terra (5.5-level at half price), but revealed the US government requested a restricted rollout instead of open access, calling it "iterative deployment." → link

• Anthropic faces mounting criticism for anti-competitive behavior and regulatory capture — Multiple developers accused Anthropic of weaponizing safety language to suppress competition, restrict access for users building alternatives, and lobbying for regulatory frameworks that entrench incumbents. → link

• Nous Research ships Mixture of Agents 2.0 in Hermes Agent — MoA presets allow combining any providers' models into virtual merged models, accessible as a single endpoint. Teknium reports significant HermesBench improvements over Opus and GPT-5.5. → link

• Ornith 35B MoE runs near-lossless agentic coding on a single DGX Spark — @sudoingX demonstrated the scaffold-RL trained model performing multi-step self-verification at FP8 (36 tok/s, 3M+ context window), highlighting that precision buys emergent reasoning behavior. → link

• DeepSeek open-sources DeepSpec speculative decoding training framework — Includes DSpark, DFlash, and Eagle3 methods; targets Qwen3 and Gemma. Community notes it's a training framework (not pre-built speedup) defaulting to 8-GPU nodes. → link

• tinygrad founder calls out US frontier lab researchers — Urged researchers to quit, arguing they used to publish but now ship only to political insiders, questioning their impact and legacy. → link

• GLM-5.2 highlighted as open-weight counter to gated US models — tinygrad reports locally hosted GLM-5.2 performs well and predicts open weights will match frontier by December. → link

• Cerebras achieves ~750 tok/s on GPT-5.6 Sol — Reported as the largest frontier model running at near-instant speeds on Cerebras hardware. → link

• OpenAI reportedly starts charging for cache writes — Hidden in fine print, signaling the end of cheap token pricing. → link

• kunchenguid's open-source projects cross 10K GitHub stars — After 3 months post-big-tech, shipped 6 projects including no-mistakes (AI slop removal), firstmate (agent), lavish (HTML editor), and AXI (agent CLI standards). → link

Analysis

Central tension — gated vs. open AI: The dominant pattern is a sharpening conflict between closed/gated frontier labs (OpenAI, Anthropic) and an open/local ecosystem (GLM, Ornith, Hermes, DeepSeek, tinygrad). Government-mandated limited access for GPT-5.6 and Anthropic's suspected regulatory capture are galvanizing the open-source community into a cohesive counter-narrative.

Local inference is quietly crossing capability thresholds: The combination of FP8 near-lossless quantization, MoE architectures (3B active / 35B total), and speculative decoding frameworks means local, single-box AI is now doing genuine multi-step agentic self-verification. This is no longer toy-grade.

Pricing pressure is shifting: OpenAI charging for cache writes and the $200/month subscription walls contrast sharply with free, locally-runnable models. The economic moat of frontier labs is under pressure from both open weights and inference cost optimization.

What to watch next: - Whether the US government limited preview for GPT-5.6 becomes a permanent access gate or a temporary measure - Speculative decode benchmarks for DeepSeek's DeepSpec at full context length - Ornith's full coding benchmark results (teased but not yet published) - Anthropic's response to sustained community backlash - Whether GLM-5.2 adoption accelerates outside the US as a result of access restrictions

Tweet Feed

AI Model Releases & Updates

@sama · 2026-06-26T20:37

Good new first: Sol is a smart, efficient, and a significant step forward. It is the same price as GPT-5.5. Also launching in the GPT-5.6 family is Terra, with 5.5-level performance at half the price. Bad news: at the request of the US government, it is launching today in limited preview instead of the open access launch we were planning on. We are working with the government to get to general availability as fast as we can. → tweet link

@sama · 2026-06-26T20:55

in other news, we updated the 5.5 instant model used in chatgpt this week. i like its vibes. → tweet link

@sama · 2026-06-26T21:06

team cooked, spicily → tweet link

@Teknium · 2026-06-26T21:07

Introducing Mixture of Agents 2.0 in Hermes Agent. Combine any provider's models into a mixture of your own. Access your presets as if it were a normal model in Hermes. Big improvement in our soon-to-release HermesBench against opus and gpt-5.5 with MoA using Opus & GPT together. → tweet link

@Teknium · 2026-06-27T06:28

Give mixture of agents a try today! → tweet link

@Teknium · 2026-06-27T18:39

RT @VaibhavSisinty: This is actually wild. Hermes just let you merge any two AI models into one virtual model. 🤯 It is called Mixture of A… → tweet link

Local AI & Open Source

@sudoingX · 2026-06-27T16:49

running Ornith on the dgx spark to see what it actually is. it's a new agentic coding model from @ornith_ / deepreinforce-ai, the 35B MoE (A3B, ~3B active per token). pulled the Q4_K_M gguf (~20GB), wired it into hermes agent, ~78 tok/s on a single spark with fast prefill, so it drives like a real agent. the part that's actually interesting is how it was trained. most coding RL just optimizes the final code. Ornith's RL optimizes the SCAFFOLD too. → tweet link

@sudoingX · 2026-06-27T17:29

i was running Ornith new 35b moe on llama.cpp with a Q4 quant… then i swapped engines. now i'm running the same MoE at FP8 in vLLM, near lossless, basically full quality, on a single dgx spark. and it's got headroom for over 3 million tokens of context on one box. → tweet link

@sudoingX · 2026-06-27T18:04

ok this is the moment the Ornith test got interesting. running the 35B at FP8… it fired a tool call to check the system. then instead of just trusting the output, it reasoned about what it got back, decided it wasn't fully sure, and called another tool to confirm. it double checked its own work before responding. self verification, mid task, on its own. the Q4 didn't do this in my runs. the precision is visibly buying you intelligence, not just cleaner text. → tweet link

@tinygrad · 2026-06-27T00:43

The US AI pay-to-play scam is so much more tolerable after switching to a locally hosted GLM-5.2. From the front page of HN, open weights will be the frontier this December. Sorry about your IPOs. → tweet link

@tinygrad · 2026-06-27T01:13

I can't believe how many people believe Anthropic's propaganda about distillation attacks. You can read how GLM-5 was trained in the paper. Maybe Claude was distilled from Chinese models? I haven't seen a paper from Anthropic, and accusations are often confessions. → tweet link

@TheAhmadOsman · 2026-06-26T22:57

This is the timeline where Opensource AI wins. → tweet link

@TheAhmadOsman · 2026-06-27T03:55

In a different lifetime I am building an Analytics company and making big money betting early. In this one I like headaches so I chose to work on making Opensource & Local AI the default. Worth it. → tweet link

@kunchenguid · 2026-06-27T04:35

woohoo! my open source projects just crossed 10,000 stars on github! very encouraging milestone right as i hit the end of 3 months after quitting my big tech job. top projects i shipped within the last 3 months: /no-mistakes, gnhf, lavish, firstmate, treehouse, AXI. → tweet link

@badlogicgames · 2026-06-26T21:38

does the us gov want to hand the international AI market to China? because that's what it looks like from afar. not sure that's super smart. → tweet link

Anthropic Controversy & Regulatory Capture

@TheAhmadOsman · 2026-06-27T17:47

Anthropic is a company wrapping a business model in moral language, then using that language to justify opaque model behavior, anti-competitive access rules, regulatory pressure, and a future where builders, startups, researchers, and Opensource communities stay downstream of a few blessed frontier labs. If a coding or research model secretly changes the quality, direction, or reliability of an answer because it classified the user as doing disallowed frontier work, the tool is no longer merely "safe." It is untrustworthy. → tweet link

@TheAhmadOsman · 2026-06-27T05:54

Anthropic is evil. Everyone should be aware of that. → tweet link

@TheAhmadOsman · 2026-06-27T09:52

Wanna know why Anthropic hates Opensource AI? GLM 5.2 being free and available to download made their $1 Trillion valuation make no sense → tweet link

@TheAhmadOsman · 2026-06-27T02:42

Five years from now when we look back we are gonna be so baffled about what a massive misstep it was to allow Anthropic to get us to this point in regulatory capture. → tweet link

@TheAhmadOsman · 2026-06-27T00:41

This is Anthropic's fearmongering dictating terms on the entire industry, and it cannot go unnoticed. Every enterprise out there needs to move as fast as possible to ensure its Intelligent Infrastructure Sovereignty. → tweet link

@tinygrad · 2026-06-27T18:55

To every researcher in US frontier labs. You used to publish. You justified not because you were shipping something millions of people love. Now you are shipping only to Trump's cronies. Consider your impact and legacy, you will be fine money wise. Monday is a good day to quit. → tweet link

@juliarturc · 2026-06-27T03:45

Sooooo are we ready to boycott Anthropic and cancel our subscriptions or are we doing the thing where we hate on them publicly but still call the Claude slot machine privately? I'm down for whatever, just let me know. → tweet link

@sudoingX · 2026-06-27T10:22

dario don't fear open weights because it's dangerous. they fear it because it's free. open weights aren't a threat to humanity. they're a threat to a valuation. → tweet link

@levelsio · 2026-06-27T03:35

I had this for about a year now. Whenever I ask Claude Code to do anything related to country codes or country names or country dropdown selectors it flags it with "Output blocked by content filtering policy." I already reported it before but they haven't fixed it, interesting. → tweet link

@gospaceport · 2026-06-26T19:09

Local AI is in the crosshairs of the government and the ChatGPT 5.6 limited release and Claude Fable pullback are probably just part of a plan 🤬 → tweet link

Inference & Hardware

@Ex0byt · 2026-06-27T14:11

Update: Single Spark+SGlang+ DeepSeek-v4-Flash (Full weights no quants). ~12.5 tok/s. I think we're just going to ship... or should we push further? → tweet link

@Ex0byt · 2026-06-27T00:41

when we bet on the right horse, great things happen. Teknium & team are killing it. → tweet link

@Ex0byt · 2026-06-27T13:24

Interesting Markov head idea (low-rank V×r transition matrix predicting next token from previous) is cheap and could in theory improve JIT expert prefetch predictors. → tweet link

@sudoingX · 2026-06-27T08:21

i read the repo before running with the hype. it's called DeepSpec, and "DSpark" is one of three methods inside it. the bigger point: it's a TRAINING framework, not a download and go speedup. you cache your target model's outputs, train a draft head yourself, then run it. no pre trained heads ship with it. still, genuinely great that deepseek open-sourced the whole spec-decode training stack. just know what it is: a framework to build the speedup, not the speedup itself. → tweet link

@alexinexxx · 2026-06-27T09:18

RT @eliebakouch: new inference optimization method by @deepseek_ai with an extremely detailed paper, draft model and framework to train the… → tweet link

@cooltechtipz · 2026-06-27T03:49

As inference costs increasingly decide what gets deployed, sparse models, from research curiosity, are becoming an economic necessity. → tweet link

@cooltechtipz · 2026-06-27T18:32

On-device AI for real-time performance. → tweet link

@cooltechtipz · 2026-06-27T14:39

AI data transfer is moving toward optical networking. → tweet link

@cooltechtipz · 2026-06-27T06:35

AI ASIC (Application Specific Integrated Circuit) vs GPU → tweet link

@cooltechtipz · 2026-06-27T08:21

How AI accelerates semiconductor development. → tweet link

Developer Tools & Infrastructure

@jack · 2026-06-27T09:21

block app kit. fastest adoption of any tool by our company. → tweet link

@sudoingX · 2026-06-27T13:25

if you're just getting into local llms, do yourself a favor and start by building llama.cpp from source. not ollama, not lm studio. build llama.cpp once, it's genuinely just a git clone and a make command with cuda on, and it clicks. → tweet link

@alexiocheema · 2026-06-26T22:59

Configuring vLLM is hard, very hard. → tweet link

@alexiocheema · 2026-06-27T17:32

Everyone is struggling, nobody wants to admit it. But I agree it's worth it when it all works. vLLM is awesome when it's all working. → tweet link

@levelsio · 2026-06-27T14:00

☁️ I made my own little Cloudflare called Pietflare, it's a DDOS and probe detector with AI and with a central IP / ASN / country block list. Each server sends suspicious probes or DDOS attempts from the access logs to the central admin and each server pulls a central blocklist every minute. → tweet link

@thdxr · 2026-06-27T04:08

ignore weird styling but we came up with a nice system to represent system prompt facts. if they change we can notify the agent without breaking the cache. this also slots in nicely with anthropic's explicit feature for this. → tweet link

@sudoingX · 2026-06-27T10:44

openclaw makes you fight the tool. hermes agent lets you fight the problem. the local ai crowd is starting to feel it. good day to upgrade your cognition tools and switch. you won't go back. → tweet link

@steipete · 2026-06-26T21:53

I love how Apple notarization breaks multiple times a year until I manually log in and accept some new legal agreements. → tweet link

@iamdevloper · 2026-06-26T23:33

just tried to read the Kubernetes docs again → tweet link

Industry & Strategy

@sudoingX · 2026-06-27T17:51

be honest, how many ai subscriptions are you paying for right now? → tweet link

@kunchenguid · 2026-06-27T16:35

i'm close to exhausting the quota from both my anthropic and openai $200/month plans, so thinking of getting a 3rd subscription. which one should i get? i care about model selection, harness flexibility and token value ROI → tweet link

@thdxr · 2026-06-27T17:38

the models are very capable. i'm not saying they're very smart. but the bottleneck with them is likely your own imagination. proof is so many companies demo their ai product with "it can schedule stuff for you" → tweet link

@nummanali · 2026-06-27T13:13

Fast intelligence is very underrated. To work effectively with agents they need to match your iteration speed. Every time you need to wait, your mind will wonder. It's the equivalent to having expert teammates on tap. → tweet link

@TheAhmadOsman · 2026-06-27T08:53

Dario would phrase it like this: "In a world full of nukes, I have a nuclear bunker." → tweet link

@gospaceport · 2026-06-27T01:16

Louie Co got a point here. How do you feel if your competition gets access to Fable class but you are stuck at Opus class? Massive disadvantage! → tweet link

@FinansowyUmysl · 2026-06-27T06:33

(Polish) I'm curious what the US government's goal is in blocking models. I'm sure it's not about security per se. Perhaps they want to somehow limit China's access. But in my opinion it will only help them. They'll gain a huge user base, and therefore data for their own training, without distillation. US companies (Anthropic, OpenAI) will have decreasing revenue growth and less training data. Seems like a terribly stupid decision. → tweet link

@FinansowyUmysl · 2026-06-26T19:13

(Polish) Amazing, the US government banned Fable 5 from Anthropic... and OpenAI just released a model of similar "power." If this continues, we'll have to switch to Codex. → tweet link

@cooltechtipz · 2026-06-27T04:52

(image link) → tweet link