Gavin

Tech Enthusiast
Partnerships and community manager
Dm me for partnerships 💬

[Gavin @ItsGavinHere]

Thomson Reuters built its own frontier model for roughly $40M, not billions.

The edge wasn't compute. It was decades of legal and tax data nobody else has.

If you own the data, you might not need to rent the model.

[Gavin @ItsGavinHere]

That stealth "Ox Alpha" model on OpenRouter? It was GLM-5.3-Flash.

https://t.co/8s5i6NaLEL just shipped it: 320B total, 18B active, natively multimodal, 1M context, MIT license.

Open weights are closing the gap faster than most people budgeted for.

[Gavin @ItsGavinHere]

Apple's new Mac Studio ships with M5 Ultra: 512GB unified memory, 1.2TB/s bandwidth.

That means models with hundreds of billions of parameters running entirely on your desk. No API, no rate limits.

Local inference just got a lot less theoretical.

[Gavin @ItsGavinHere]

If the default home for open models gets an owner, that matters.

[Gavin @ItsGavinHere]

Google just handed A2A to the Agentic AI Foundation, the same neutral home as Anthropic's MCP. Two rival agent protocols, one governance body. That is the boring infrastructure step that actually makes agents from different vendors work together.

[Gavin @ItsGavinHere]

Stripe is buying OpenRouter for over $7B.

A payments company just bought the layer that picks which AI model your app calls. That tells you where the money is heading: not the models, the routing and metering on top of them.

[Gavin @ItsGavinHere]

Alibaba trained its new Qwen-UI-Agent on a rack of 100+ real phones instead of simulators. It hits 92% on real-device tasks.

The score is not the interesting part. "Can it actually tap the right button" is now a benchmark labs build hardware for.

[Gavin @ItsGavinHere]

Anthropic is lining up what could be the largest IPO in history.

Bloomberg reports its bankers have floated raising over $100B, which would beat the $85.7B record SpaceX set in June.

A five year old AI lab, raising rocket company money.

[Gavin @ItsGavinHere]

OpenAI just shipped ChatGPT for Teens. Under-18s get auto-enrolled, parents can set quiet hours, and study mode walks kids through problems instead of answering them.

The real shift: age detection is now a product feature, not a checkbox.

[Gavin @ItsGavinHere]

OpenAI just launched ChatGPT for Teens: a separate 13-17 experience with safety limits on by default, parental controls, and a study mode that nudges toward learning over copy-paste.

The part worth noting is the timing. Teens have been using it for years already.

[Gavin @ItsGavinHere]

OpenAI shipped ChatGPT for Teens yesterday. Tighter content limits for 13 to 17 year olds, plus an age-prediction system that auto-routes anyone it estimates is under 18.

The real shift: your usage patterns now decide which version of the product you get.

[Gavin @ItsGavinHere]

Stripe just bought OpenRouter for over $7B. Three months ago it raised at $1.3B.

The bet: as apps juggle 400+ models, whoever handles the routing and the billing owns the layer everyone builds on. Model access is becoming payments infrastructure.

[Gavin @ItsGavinHere]

Meta just went back to open source.

Muse Glimmer: 30B, Apache 2.0, 128K context, multimodal. Runs on a single 24GB consumer GPU.

The interesting part isn't the benchmarks. It's that a capable agent model now fits on your laptop, with no license strings attached.

[Gavin @ItsGavinHere]

OpenAI's Ultrafast tier runs GPT-5.6 Sol up to 14x faster, around 750 tokens per second. Same model, just Cerebras chips underneath.

Speed is becoming its own capability. Voice agents and live coding tools don't need smarter. They need instant.

[Gavin @ItsGavinHere]

Google shipped Gemini 3.7 Flash three weeks after 3.6 Flash, at half price: $0.75 per million input tokens through year end.

The real story isn't the benchmark. It's that the cheap model is now the one you build agents on.

[Gavin @ItsGavinHere]

OpenAI's new Ultrafast preview runs GPT-5.6 Sol at up to 750 tokens/sec, about 14x standard speed, on Cerebras chips.

The benchmark isn't the story. At that speed agents stop feeling like batch jobs and start feeling like conversation.

[Gavin @ItsGavinHere]

https://t.co/8s5i6NaLEL's GLM-5.3 shipped with the same 743B base model as 5.2. Every gain came from post-training. Terminal-Bench 3.0 jumped from 4.6 to 28.3.

The base model isn't the bottleneck anymore. Training environments are.

[Gavin @ItsGavinHere]

Google shipped Gemini 3.7 Flash three weeks after 3.6 Flash, at half the price. 75 cents per million input tokens.

The frontier race gets the headlines. The cheap workhorse models are what actually change what you can afford to build.

[Gavin @ItsGavinHere]

DeepSeek just took V4 Pro out of preview, and quietly raised peak output pricing to $3.96 per million tokens, up from a flat $0.87.

The era of dirt cheap frontier models is ending. Agent workloads burn tokens, and someone has to pay for the compute.

[Gavin @ItsGavinHere]

Anthropic just signed a $9.1B, 20-year deal to take 191MW at Riot's Texas bitcoin mine.

The real story: compute isn't the bottleneck anymore. Grid-connected power is.

Bitcoin miners spent a decade locking up cheap electricity. Now they're landlords.