AI news in context

ChatGPT is about to get a voice mode upgrade as a new “gpt-bidi-1” model has been spotted along with announcement updates. Community angle: discussion is focused on practical implications and community skepticism

Anthropic is a predatory company. Anthropic's support and refund system is built to take your money and make sure you never see it again. I'm a young adult. I scraped money together to use Claude. Three days before their newest model came out I paid $125 for Max 5x. When the new model dropped, I paid another $200 for Max 20x on top of that so I can use fable for longer than 30 minutes.. Three days later the model was gone. I understand that it was a government export control order. What I am angry about is Anthropics "support system" I went to the support chatbot to request a refund. My message was about the Max 5x and Max 20x subscriptions. T. Community angle: discussion is focused on practical implications and community skepticism

Glimmer 1 - Glint Research. A foundational 10,000 parameter language model. We introduce Glimmer, a 10k base model trained on 500K tokens of FineWeb-Edu. The context window is 512 tokens The arch is standard llama (LlamaForCausalLM) 16 hidden dims 2 layers 4 attention heads 1 KV head (GQA) And the rest is on AMA for as long as this post is up. Have fun! \# Benchmarks * arc\_easy (acc): 25.46% * wikitext-2 (word\_perplexity): 1,765,201 * wikitext-2 (byte\_perplexity): 14.73 * wikitext-2 (bits\_per\_byte): 3.8806 * BLiMP (acc): 52.43%. Community angle: discussion is centered on benchmark quality

after a few weeks I've basically settled into a 3-model split and I'm curious how different everyone else's is. I used to just run Opus on everything because "best model, why not." that was dumb and expensive in terms of hitting limits. where I landed: Opus 4.8 - anything where being wrong is costly. real code, tricky reasoning, anything I'm going to ship or send. it's noticeably better at not letting its own mistakes through, which matters most exactly here. Sonnet 4.6 - the actual workhorse. drafting, rewriting, most of my thinking-out-loud, first passes on basically everything. fast enough that I don't context-switch waiting, good enough that I rarely need to escalate. Haiku 4.5 - the cheap fast lane. reformatting, q. Community angle: discussion is centered on developer workflow

possible Fable comeback?. pinging Fable of azure foundry used to yield the same response as the model is unavailable and suggesting a switch to opus 4.8. For the last couple of hours, Anthropic models have been a bit iffy on the api side on foundry and when i pinged Fable just now, i got this. Fingers crossed. Community angle: discussion is centered on developer workflow, privacy and data use, and infrastructure costs

How many tabs do you currently have open just to work with AI?. Mine usually looks like: * ChatGPT * Claude * Gemini * Perplexity * Google Docs And that's before I actually start working. Has anyone found a workflow that avoids constantly switching between tools?. Community angle: discussion is focused on practical implications and community skepticism

Performance nose-dive. Unusable. Are these companies doing something to the models the week prior to a new release? Seriously, I've seen this same pattern with Claude in the past and now Codex hit a session today that is the same. I'm just learning about what a quantized model is and what can happen when you use one. Is this what they're doing to their own models to make them basically stroke out? This is a list of all the problems I've had in the last hour and a half, none of which I've had happen with Codex before. It's completely ignoring the system we've been using. EDIT: Hermes seems to be able to avoid the stoked out part of Codex, for. Community angle: discussion is centered on reliability

LOOP.md: a Markdown-native while loop for Codex-style agents. I kept running into the same gap when using Codex and other coding agents: Agents are good at doing a task once, but “continue until X” usually ends up buried in a long prompt, hidden session state, or tool-specific behavior. So I made the loop explicit in Markdown: SKILL.md: Do this once. LOOP.md: Do this again when triggered, carrying state and record from prior iterations. Minimal file model: LOOP.md = while-loop definition state.md = current loop variables / status record.md = append-only iteration ledger Core rules: Every activation must be bounded. The bound does not have to be numeric. A loop does not r. Community angle: discussion is focused on practical implications and community skepticism

I built a deterministic drive tracking algo on iOS, in partnership w/ Claude Code. I'm Josh, a solo dev. About a month ago I shipped EveryLastMile — an iOS mileage tracker — to the App Store, built nights and weekends with heavy use of Claude Code. It's absolutely free to try (30-day trial, no credit card). I wanted to write up where Claude Code genuinely carried the work and where it hit a wall, because the wall taught me more than the wins did. **The meat of my iOS build** The hard part of a mileage tracker is detecting a drive in the background without cooking your battery, and locking the drive's origin before GPS catches up to where you actually started. I built it on Swift 6.2 with str. Community angle: discussion is focused on practical implications and community skepticism