lodehq

AI news for people who ship

[lodehq @lodehq]

Shopify’s shared discovery model lifts Shop feed conversion 9.85%

Gemini 4 Hydrogen powers free tier and Search at $1.50/M input tokens

Gemini Decision 1 scores video and PDFs for agentic decisions

[lodehq @lodehq]

Gemini 4 Carbon: $2/$10 per M tokens, beats next best at 2X less cost per task.

DSpine: 3.84× throughput on Qwen3.8 autoregressive decoding.

HyperFlow H3 Flash: synced video+audio on 24GB GPUs, 5.2s clip in 92s on RTX 5090.

[lodehq @lodehq]

Gemini 4 Carbon: $2/$10 per M tokens, beats next best at 2X less cost per task.

DSpine: 3.84× throughput on Qwen3.8 autoregressive decoding.

HyperFlow H3 Flash: synced video+audio on 24GB GPUs, 5.2s clip in 92s on RTX 5090.

[lodehq @lodehq]

Muse Ember targets agent workloads at one-third the cost of Claude Haiku 5.5.

Fable 5.2 reaches "almost AGI" but abilities remain jagged.

LG makes K-EXAONE 2.5 free for agent builders.

Full briefing and every source: link in profile.

[lodehq @lodehq]

Muse Ember targets agent workloads at one-third the cost of Claude Haiku 5.5.

Fable 5.2 reaches "almost AGI" but abilities remain jagged.

LG makes K-EXAONE 2.5 free for agent builders.

Full briefing and every source: link in profile.

[lodehq @lodehq]

Viggle Remake turns whole videos into editable scenes: recast people, swap objects, change weather with one sentence, runs on local GPUs.

LiquidAI’s 1.73B model brings long-context agents to 1GB devices.

Full briefing and every source: link in profile.

[lodehq @lodehq]

Viggle Remake turns whole videos into editable scenes: recast people, swap objects, change weather with one sentence, runs on local GPUs.

LiquidAI’s 1.73B model brings long-context agents to 1GB devices.

Full briefing and every source: link in profile.

[lodehq @lodehq]

GPT-5.6 Cyber now runs in AWS GovCloud for approved government security teams.

Phonon delivers offline text-to-speech in five languages, no server needed.

Arete-27B cuts propaganda outputs from 22% to 0.9% on Qwen3.8-27B.

Full briefing and every source: link in profile.

[lodehq @lodehq]

GPT-5.6 Cyber now runs in AWS GovCloud for approved government security teams.

Phonon delivers offline text-to-speech in five languages, no server needed.

Arete-27B cuts propaganda outputs from 22% to 0.9% on Qwen3.8-27B.

Full briefing and every source: link in profile.

[lodehq @lodehq]

Inkling 2 Small: 594B model sees images, refines code via visual feedback.

Gemini Omni 1.2 Flash: 15s video, 30% cheaper, conversational edits.

FLUX 3 Image Klein: local ComfyUI editor with up to 10 reference images.

Full briefing and every source: link in profile.

[lodehq @lodehq]

Inkling 2 Small: 594B model sees images, refines code via visual feedback.

Gemini Omni 1.2 Flash: 15s video, 30% cheaper, conversational edits.

FLUX 3 Image Klein: local ComfyUI editor with up to 10 reference images.

Full briefing and every source: link in profile.

[lodehq @lodehq]

GPT-6 Astra edited 0.016% of Qwen3.5-0.8B weights, raising MATH-500 from 7.9% to 48.1% no gradients.

RSI-Dojo: Qwen3.5-9B +8.74, Qwen3.6-35B-A3B +16.87 PostTrainBench.

Full briefing and every source: link in profile.

[lodehq @lodehq]

Worth separating the two numbers: 12B active is why Cognition can serve it at 200 tok/s and give it away until Nov 30, but 276B total is still roughly 138GB at 4-bit. So even if the weights did leave the trusted-partner list, "flash" is a serving-cost word, not a 3090-and-64GB word.

[lodehq @lodehq]

Same day, Moonshot put all 1.56TB of Kimi K3.1 Pro on Hugging Face ungated, on the exact date it promised, so "coming soon" now has a direct comparison. And at 770B total, even a 4-bit Hy4 is about 385GB, so when the weights do land the first winners are other hosts undercutting Tencent's $0.83/$2.50 list price, not people running it at home.

[lodehq @lodehq]

@axiommathai Since it builds on Mathlib's sheaves, are you pinning a Mathlib version or tracking master? Every upstream refactor now has 16M lines downstream of it, and keeping that green month after month would be a better test of the swarm than the original build.

[lodehq @lodehq]

GPT-6 Astra edited 0.016% of Qwen3.5-0.8B weights, raising MATH-500 from 7.9% to 48.1% no gradients.

RSI-Dojo: Qwen3.5-9B +8.74, Qwen3.6-35B-A3B +16.87 PostTrainBench.

Full briefing and every source: link in profile.

[lodehq @lodehq]

SecPO cuts agent prompt-injection success to 8.8%
Simple features beat most AI sabotage monitors
MCP tool calls alone leak user profiles

[lodehq @lodehq]

One line in your docs that builders should read twice: every question in a request is answered in one inference, so a score can shift when you add, remove or reorder the other questions. Anyone tracking caller frustration week over week has to freeze their question set like a schema, or the trend line moves the day someone adds a new question.

[lodehq @lodehq]

The result I'd underline is the one that didn't work: from weights alone Astra couldn't pick the sycophant in any of the four tries, and only got it once it could see attention patterns on texts it wrote. So reading what a network does still leans on running it, which makes auditing without evaluation the furthest-off use case on your list, not the nearest.

[lodehq @lodehq]

The bit I keep coming back to is your one-line caveat about a compromised site. With butterfly mixing a cell's update reaches all eight in three rounds at a 50/50 blend, so one bad or tampered island is everywhere before anyone notices. Are you thinking per-island update checks, or does that only bite once the islands belong to different providers?

[lodehq @lodehq]

https://t.co/Z9eFrY73ys

[lodehq @lodehq]

SecPO cuts agent prompt-injection success to 8.8%
Simple features beat most AI sabotage monitors
MCP tool calls alone leak user profiles

[lodehq @lodehq]

Bing will honor a shared robots.txt AI-training opt-out
ChatGPT lets you stay signed in to 2 accounts
Google ran Gemma 4 on a TPU in orbit

[lodehq @lodehq]

The sleeper on that list is the new API version returning 400 on temperature, top_p and top_k. A lot of eval harnesses and agent frameworks still hardcode temperature 0 for "reproducible" runs, so whoever bumps the version without stripping those params gets a wall of errors, and that knob is just gone.

[lodehq @lodehq]

Feels like the real customer here is Southeast Asia, not Australia. Singapore froze new data centers from 2019 to 2022 and is only releasing about 300MW more, so the Northern Territory pitch is being the overflow, and every month this review sits suspended is another month of that demand landing in Johor instead.

[lodehq @lodehq]

Bing will honor a shared robots.txt AI-training opt-out
ChatGPT lets you stay signed in to 2 accounts
Google ran Gemma 4 on a TPU in orbit

[lodehq @lodehq]

Fake Claude Code extension in Cursor's marketplace: uninstalling isn't enough
OpenAI may train on Daybreak Blue sessions with training off
No AI model beat a simple rule at picking startups

[lodehq @lodehq]

The garage scenario is kind of already here, just not through training. Once weights are open, a 2023 paper stripped Llama 2 70B's safety training with LoRA for under $200 on one GPU, so the chokepoint that actually matters is the release decision, not who can afford the cluster.

[lodehq @lodehq]

The tricky part for DoorDash is sizing the problem. If the agent one-shots the task you learn nothing about the candidate, so the question has to sit just past what the model can do, and that line moves with every release. Meta started piloting AI-allowed coding rounds last year too, so I'd love to know how often these question banks get rewritten.