This is a super exciting release - Claude Fable 5 is the same underlying model as Mythos but with added safeguards. The benchmarks are great and it's SOTA on everything by a margin but I'll add that *qualitatively* also, this is a major-version-bump-deserving step change forward (imo of the same order as Claude 4.5 was in November), peaking especially for long problem-solving sessions on very difficult problems. You can give it a lot more ambitious tasks than what you're used to, the model "gets it" and it will just go, and it's never felt this tempting to stop looking at the code at all (but don't do this in prod!). The model still has quirks that people will run into and the safeguards are configured to be a little too trigger happy for launch, which can hopefully be tuned over time. I feel a lot of things changing as working software increasingly comes out on a tap. The Jevon's paradox kicks in and I feel my own demand for software growing substantially. You can ask for anything - explainers, visualizers, dashboards, bespoke single-use apps (e.g. a full wandb that is hyper-specific just for your project), you can 10X your test suite, auto-optimize code, run giant research projects with custom HTML for the results, anything! "Free your mind" (Matrix ref). Really looking forward to all the things people build!

Okay, GPT 5.5 Instant is AMAZING They absolutely nailed it on the personality! Well done guys 👏👏 @sama @gdb @OpenAI

New voice mode: - powered by GPT-5.5 instant instead of 4o as it is right now (possibly the biggest single update in chatgpts history) - BiDi: it can talk while listening and listen while talking. You can talk over each other, the conversations wont be turn based anymore, and this makes it 100x better and more natural and more intelligent. Imagine you are yapping something and it says "yeah" or "oh I see" while you talk, or when it yaps you can say or ask something and it dynamically reacts to it, basically just how human-human conversations are - even normal people who dont care much about AI will recognize that this is a huge upgrade on all 3 axes, personality, intelligence and immersion - probably comes within the next two weeks, possibly tomorrow - I guess this will be one of the top 3 moments in AI of 2026 for me, possibly bigger than images v2 which i waited a year for - it might be able to do not just talk, but also modulate its voice to whisper or yell, or to express different feelings, laugh, sing and so on, 4o could do all that - hope it wont get nerfed again

LLM Knowledge Bases Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge (stored as markdown and images). The latest LLMs are quite good at it. So: Data ingest: I index source documents (articles, papers, repos, datasets, images, etc.) into a raw/ directory, then I use an LLM to incrementally "compile" a wiki, which is just a collection of .md files in a directory structure. The wiki includes summaries of all the data in raw/, backlinks, and then it categorizes data into concepts, writes articles for them, and links them all. To convert web articles into .md files I like to use the Obsidian Web Clipper extension, and then I also use a hotkey to download all the related images to local so that my LLM can easily reference them. IDE: I use Obsidian as the IDE "frontend" where I can view the raw data, the the compiled wiki, and the derived visualizations. Important to note that the LLM writes and maintains all of the data of the wiki, I rarely touch it directly. I've played with a few Obsidian plugins to render and view data in other ways (e.g. Marp for slides). Q&A: Where things get interesting is that once your wiki is big enough (e.g. mine on some recent research is ~100 articles and ~400K words), you can ask your LLM agent all kinds of complex questions against the wiki, and it will go off, research the answers, etc. I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this ~small scale. Output: Instead of getting answers in text/terminal, I like to have it render markdown files for me, or slide shows (Marp format), or matplotlib images, all of which I then view again in Obsidian. You can imagine many other visual output formats depending on the query. Often, I end up "filing" the outputs back into the wiki to enhance it for further queries. So my own explorations and queries always "add up" in the knowledge base. Linting: I've run some LLM "health checks" over the wiki to e.g. find inconsistent data, impute missing data (with web searchers), find interesting connections for new article candidates, etc., to incrementally clean up the wiki and enhance its overall data integrity. The LLMs are quite good at suggesting further questions to ask and look into. Extra tools: I find myself developing additional tools to process the data, e.g. I vibe coded a small and naive search engine over the wiki, which I both use directly (in a web ui), but more often I want to hand it off to an LLM via CLI as a tool for larger queries. Further explorations: As the repo grows, the natural desire is to also think about synthetic data generation finetuning to have your LLM "know" the data in its weights instead of just context windows. TLDR: raw data from a given number of sources is collected, then compiled by an LLM into a .md wiki, then operated on by various CLIs by the LLM to do Q&A and to incrementally enhance the wiki, and all of it viewable in Obsidian. You rarely ever write or edit the wiki manually, it's the domain of the LLM. I think there is room here for an incredible new product instead of a hacky collection of scripts.

I built Feynman, Claude Code for research. I gave it a question and it came back 30 minutes later with a cited meta analysis. It can also replicate experiments on Runpod, audit claims against code, and simulate peer review. Open source & MIT license, link below

This makes more sense as a workflow feature than an “agents” feature. Parallelizing pieces of a task is nice, but managing context without clutter is the real unlock.

The MacBook Air M5 battery is so good it’s actually messing with my habits. I keep reaching for the charger out of instinct. Then realise I don’t need it. 🤣🤣

ARC is still a weird benchmark, but the efficiency trend is hard to ignore. Reasoning is slowly moving from “expensive demo” to “cheap enough to use everywhere.

This is probably the most compelling version of “vibe coding” so far: not shipping a startup, just building the exact software you wish existed for your own life

alpha-gpt-5.4 was just spotted in a public models endpoint and those deleted codex github pr leaks are 100% real. openai is gearing up to drop this way sooner than anyone anticipated, likely as early as next week, and it is going to completely reset the board. forget the incremental updates we have been getting lately; this is the true generational leap we have been waiting for, and here is exactly why it is going to melt your mind.

I stopped paying for Claude Code. $200/month for an API subscription to write code. Then Ollama dropped Anthropic API compatibility. Now Claude Code connects to free, local models on my machine. Here's the exact setup (took me 10 minutes): 1. Install Ollama → curl -fsSL https:// ollama .com/install .sh | sh 2. Pull a model → ollama pull qwen2.5-coder 3. Point Claude Code at localhost → ANTHROPIC_BASE_URL=http: //localhost :11434 claude That's it. Claude Code thinks it's talking to Anthropic. It's talking to your laptop. Best models I've tested: • Qwen 2.5 Coder - best all-around for code generation • DeepSeek-Coder - strongest at debugging and refactoring • Llama 3 - solid general reasoning Local models aren't Sonnet or Opus. Complex multi-file refactors still stumble. Long context windows get messy. But for everyday coding - scaffolding, tests, quick edits, boilerplate - they handle it fine. Your code never leaves your machine. Your bill goes from $200/month to $0. Your API key stays in your pocket. (Save this for later.)

This aired tonight to 1 billion people in China. A year ago these robots could barely wave a handkerchief, now they can do backflips and kung fu with nunchucks. Physical intelligence is the next frontier.

Codex and Claude Code are now so good at assembling long research reports that I no longer need to browse sources on the web myself. If you want to compile a topical report with references, links, open questions, simulations, and so on, you're better off creating a list of skills for your favorite agentic tool before you even start looking at the sources. I've been testing both Claude Code with Octopus and Codex-Spark (JupyterBook skill I wrote myself) for writing very long website/PDF reports (around 150-200 pages) on specific topics, complete with references. It takes about 3-6 hours (with light supervision and occasional inspection) to produce a very comprehensive guide to a given subject. Why would I browse the web directly myself? I can write an agentic skill and obtain something tailored precisely to my needs. This is the future of web interaction. You build scaffolds around the raw web instead of relying on traditional browser search. Want to check the original source? Sure, but you do it in the context of your report. And it's becoming incredibly fast. It also makes much more sense to maintain raw markdown files for this kind of report generation rather than writing fixed booklets for posterity. Knowledge becomes a graph of relations (with nodes as tiny portions of insight), and you maintain this relational network, then reconstruct a specific report on demand.

gpt-5.3-codex for rewriting applications between languages:

Google DeepMind 🤝 @BostonDynamics Our new research partnership will bring together our advancements in Gemini Robotics’s foundational capabilities to their new Atlas® humanoids. 🦾 Find out more → goo.gle/49paguA

There is no neutral way to say this: Donald Trump is siding with the enemy of Europe and the United States. The Soviet Union was built to destroy the Western world. It was an authoritarian system rooted in control, repression, and the elimination of individual freedom. It aimed its nuclear arsenal at America and trained generations to believe that democracy was weak, corrupt, and destined to fall. That system collapsed, but its mindset did not. It survived in Moscow, reshaped itself, and waited. Today’s Russia is not communist in name, but it is Soviet in instinct. Power over law. Loyalty over truth. Strength over freedom. And above all, hostility toward the Western democratic order. This is the enemy the West was built to resist. And yet Donald Trump openly aligns himself with it. He attacks NATO while praising authoritarian leaders. He questions democratic elections while echoing Kremlin narratives. He undermines Ukraine while flattering the regime that invaded it. Trump does not support Russia because he loves Russia. He supports it because he shares its contempt for democracy. He admires systems where power is personal, institutions are weak, and loyalty matters more than law. In that world, Putin is not a threat. He is a model. That is what makes this moment so dangerous. Not Russian strength, but Western self-sabotage. Not tanks or missiles, but the normalization of authoritarian thinking inside democratic societies. Europe cannot afford to misunderstand this. This is not about left or right, or policy disagreements, or personality clashes. This is about allegiance. And Trump has chosen his side.

To all the Ukrainian warriors in trenches tonight on Christmas defending freedom in Ukraine, Europe, and the world, thinking of you. In awe of you. #SlavaUkraini !

It is a very smart model, and we have come a long way since GPT-5.1: