Yesterday I said Jev would open a ton of doors... 24 hours later, this exists. Cua built a 2.8MB model that scored 99.7% on their form-filling eval. Hosted Jev scored 83.6%. Not to mention it's FREE and only 706K parameters. Small enough to run locally with not even 1gb or ram. Fast enough to make decisions in one pass. And specialized enough that your agent doesn't need to call a giant LLM for every tiny action. Think about what this unlocks. Every repetitive computer task could eventually get its own tiny specialist: • forms • CRM updates • data entry • browser actions • document routing • UI decisions Then one powerful agent just routes work between them. We are going to see some ridiculous stuff built from this as well.

Imagine if subscriptions didn't exist... ONE of my Claude Code sessions today: > $200.66 API-equivalent cost > 98% Opus 5.5 > 98% cache hit And Opus 5.5's cached input is 20x cheaper: > normal input: $4/M tokens > cached input: $0.20/M > output: $20/M So almost the entire session ran at the discount rate... and it STILL came out to ~$200. How the heck do companies pay API prices for models like Opus 5.5?

Opus 5.5 has NOT been nerfed.... Nerf Bench has tested it 7 times since launch. > launch: 100% > lowest: 94.2% > latest: 105.8% i swear people are just making this up for engagement. not to mention, even if it was 5% worse, you wouldn't be able to tell anyway. and of course it isn't.

the fastest way to make a cold DM feel automated is trying too hard to personalize it. most AI outreach tools do the same thing: > scrape someone's profile > find one random detail > force it into the opener > call that personalization Leadpoet goes way deeper. it looks through: > job listings > LinkedIn > 𝕏 > the open web and tries to find an actual reason that person might need what you're selling right now. so instead of: "loved your post... anyway, here's my pitch" you can reach out because their company is hiring for something your product solves, they're publicly complaining about the exact problem, or there's some other actual buying signal. automate the research that gives you a reason to reach out in the first place.

Sam Altman watching DeepSeek slowly open-source the entire AI stack: today they open-sourced DeepInfer, the framework they use to run V4.1 Flash at scale. you get: > optimized inference kernels > distributed inference > speculative decoding > async scheduling > structured outputs > NVIDIA + Ascend support and it's Apache 2.0. they've already released model weights, research, training recipes, their coding harness, and now the inference framework running V4.1 Flash. every few months another layer of DeepSeek becomes public. at this rate they're going to open-source the whole stack piece by piece.

Claude loves releasing models while I’m golfing. Opus 4.8 came out when I was golfing. Opus 5 came out when I was golfing. Fable 5.1 came out when I was golfing. Opus 5.5 came out when I was golfing. Well I’m golfing again tomorrow… So I wouldn’t be surprised if Fable 5.5 came out tomorrow lol.

i think people are making building software WAY harder than it needs to be. there is already software for basically everything. i turned Hermes Bot Mode into my own desktop AI agent instead of rebuilding it from scratch. here's how i'd build almost anything now: 1. FIND A PROBLEM PEOPLE ALREADY PAY TO SOLVE don't ever start with “what can i build?” i've done that plenty of times. it never went anywhere. find a workflow a business already spends money or employee hours on, then figure out who owns it, what they do today, where time gets wasted, what software is involved, and what fixing it is worth. keep the first customer painfully specific. “software for contractors” tells you nothing. “take roofing leads, qualify them, create an estimate, follow up, and push everything into their CRM” gives you a workflow to build around. 2. SEARCH FOR WHAT ALREADY EXISTS before writing code, search GitHub like you're trying NOT to build the whole thing yourself. ChatGPT deep research is great for this. Give it the workflow and ask for open-source products, libraries, APIs, abandoned projects, and software from adjacent industries. compare the best options by: > license > maintenance > tech stack > deployment > integrations > how much of the workflow already exists > how hard it'll be to modify pick whatever removes the most work between where you are and getting customers. 3. UNDERSTAND IT BEFORE YOU CHANGE IT give the repo to Claude Code and have it map the entry points, important folders, data flow, schema, auth, APIs, background jobs, external services, tests, and where your workflow lives. then ask: “If I wanted to turn this into [your exact product], what should stay, what should change, what should be deleted, and what needs to be built from scratch?” if you need to understand software beyond normal source code, REA (Reverse Engineer Anything) can help inspect how the application works. get a modification plan before Claude starts editing. 4. DELETE BEFORE YOU ADD the repo probably has WAY more than your customer needs. remove features they won't use, settings they shouldn't need, irrelevant integrations, extra navigation, dead dependencies, and workflows that don't apply. every extra feature is another thing to test, explain, secure, maintain, and keep from breaking. i'd rather cut down something powerful than rebuild auth, databases, permissions, APIs, integrations, and everything else from zero. 5. REBUILD IT AROUND ONE BUSINESS map the workflow from start to finish. for a roofing company: lead comes in → qualify → estimate → send → follow up → update CRM → alert salesperson then make the software follow that flow using the tools they already have: CRM, email, forms, calendar, database, payments. the software should fit how the business already works. 6. ONLY BUILD THE GAP after that, the amount of NEW software you need can be surprisingly small. maybe it's: > one integration > one automation > one dashboard > one agent > one approval flow > one weird industry-specific feature build it, put it in front of the customer, and watch what they ignore, what confuses them, what they still do manually, and what they immediately ask you to change. that tells you what to build next. 7. SELL WHAT IT DOES FOR THEM the customer doesn't care what repo you started from or how much code Claude wrote. they care that: > leads get followed up > reports stop taking 4 hours > quotes go out faster > data stops getting copied by hand > fewer customers fall through the cracks > employees get hours back charge for solving that. then when another business in the same niche has the same problem, most of the system is already built. you usually don't need some genius new technology to make useful software. most of the hard parts probably already exist for free. your job is to find them, cut them down, and build the part that specific customer is still missing.

this is getting stupid lol DeepSeek V4.1 Flash was already free on OpenCode. the better 1017 checkpoint drops... and they just made THAT free too. how is anyone supposed to compete with this. fun fact: they do make money after all... a lot of money.

Fable 5.5 could FINALLY drop this week... quick rundown of everything we actually know / that's leaked so far: > rumored to be a fresh pretraining run, meaning NEW base model weights rather than Fable 5.1 with another round of post-training > `claude-fable-5-5` briefly appeared in Azure's model catalog 5 days ago, then disappeared > partner access has reportedly been circulating for the past few days > leaked Azure metadata showed 1M context + 128K max output > that same metadata showed $10/M input + $50/M output > BUT those context + pricing numbers appear to have been copied from Fable 5.1, so they could just be placeholders > Anthropic has confirmed newer Claude models are increasingly being built with Claude itself, including writing code + running research experiments > if Anthropic keeps the same hierarchy, Fable 5.5 should sit ABOVE Opus 5.5 as their most capable model a completely new base model would be insane... if that leak is right, Fable 5.5 should be a MUCH bigger jump than another 5.x checkpoint.

translation: Anthropic thinks Fable 5.5 is about to print so much money they're already building the system to redistribute some of it... before the model is even released lol

i've gotten WAY too good at switching models around my weekly limits... i've been HAMMERING them all week and somehow i still have a ton of usage left. but it resets today. so now every repo is open, every agent is working, and every stupid experiment i've been saving is getting launched at once lol Opus 5.5 is genuinely too good at not burning my usage limits.

OpenAI just made an 8-month-old mystery WAY more suspicious... back in February, a completely anonymous model called Aurora Alpha appeared on OpenRouter. all we knew: > 128K context > reasoning model built for speed > coding assistants > real-time applications > agentic workflows 3 days later, OpenAI released Codex-Spark. also: > 128K context > built for insane speed > real-time coding > agentic software work people immediately wondered if Aurora Alpha was an early Spark model. but it was NEVER revealed. now 8 months later, three models suddenly appear in OpenAI's own API: > aurora-red-alpha > aurora-green-alpha > aurora-blue-alpha same Aurora name. same Alpha naming. that old model being OpenAI suddenly feels a LOT less random... now what the heck are red, green and blue???

I tested 3 repo-map tools to fix one of Claude Code's dumbest mistakes: it'll make the change you asked for, then miss other places in the repo that need to change with it. after testing all 3, the best fix wasn't a repo-map tool at all. it was one sentence in CLAUDE.md. > "Before changing a function, data shape or stored value, search the whole repo for every place that calls, reads or writes it, and update them all." on a fresh set of tasks: > plain Claude Code: 18/24 > with that one sentence: 24/24 token use was basically the same. then on another held-out set, where I injected the repo maps at the start: > plain: 9/12 > Tessera pushed: 9/12 > repomap pushed: 9/12 > one sentence: 11/12 when I only made the repo-map tools available instead of forcing them into context, Claude barely used them: > pi-shazam: 0/54 > Tessera: 0/54 > repomap-mcp: 22/54 the one miss on the held-out set explained why the sentence could beat a repo map. a CLI edited the data file directly without ever calling the function. the repo map followed the code relationships. the instruction made Claude search for every read + write across the repo. small eval, and I deliberately made the tasks contain this kind of failure, so I'm not saying repo maps are useless. but before testing this, I would've absolutely assumed the extra tooling would help more. one sentence in CLAUDE.md did better.

if my entire Claude Code setup got wiped today, here's exactly how i'd rebuild it in 30 minutes: 1. ROUTE THE MODELS FIRST i wouldn't use the same model for everything. Sonnet 5.5: > well-scoped feature work > tests > UI work > repetitive implementation > repo exploration Opus 5.5: > bug fixes > code review > architecture > big refactors > ambiguous changes especially for bugs, i want Opus tracing the problem BEFORE it edits anything: reproduce → trace the call/data flow → prove the root cause → make the smallest fix → add the regression test. and i'd always rather have a fresh Opus session review important code than the same session that wrote it. use /model + /effort instead of leaving one model on everything. 2. KEEP CLAUDE.md SMALL run /init, then clean up what it generates. my root CLAUDE.md would only have stuff Claude should know constantly: > what the project does > important architecture > build/test/lint commands > core conventions > areas it shouldn't casually touch then i'd put specific rules in .claude/rules/. frontend rules only load around frontend code. database rules only matter around database code. test rules only matter around tests. there's no reason to make Claude reread 500 lines of instructions on every task. 3. TURN ON CODE INTELLIGENCE this is one i see so many people skip. install the Claude Code code-intelligence/LSP plugin for whatever language you're using. now Claude can work with: > symbol definitions > references > types > diagnostics > errors immediately after edits instead of constantly searching strings and opening half the repo trying to figure out where something comes from. on a big typed codebase this is a HUGE upgrade. 4. MAKE SKILLS FOR HOW YOU ACTUALLY CODE i'd start with: /investigate /debug /review /ship /investigate: map the relevant code, trace the flow, find existing patterns + tests, but DON'T edit anything. /debug: reproduce the bug, build hypotheses, test them, prove the root cause, then make the smallest fix. /review: read the diff, trace what it affects, look for regressions, bad assumptions, scope creep + missing tests. /ship: check the diff, simplify anything unnecessary, run validation, then give me exactly what changed. if i type the same 5-step prompt more than twice, it becomes a skill. 5. USE SUBAGENTS TO KEEP THE MAIN CONTEXT CLEAN i don't want my main Claude session reading 80 files just to answer one question. i'd have narrow agents for: > repo research > Opus code review > test/edge-case hunting the research agent can disappear into the codebase and only send the important findings back. the reviewer gets fresh context instead of inheriting all the assumptions from the builder. and if i want 3 different opinions on a change, run them in parallel instead of making one Claude pretend to review itself three times. 6. PUT GUARANTEES IN HOOKS, NOT PROMPTS if something NEEDS to happen every time, i don't want to hope Claude remembers it. hooks should handle stuff like: > format edited files > run lint after edits > block edits to protected files > stop dangerous commands > run typecheck/tests before finishing same thing with permissions. /permissions should auto-allow the boring safe stuff: > reads > searches > git diff > tests > lint > typecheck and keep approval around: > deploys > production > migrations > secrets > destructive commands CLAUDE.md tells Claude what you want. hooks + permissions can actually enforce it. 7. STOP DESTROYING YOUR CONTEXT WINDOW /context is underrated. use it and actually look at what is eating your context. then: > /compact when a long task starts bloating > /clear before an unrelated task > /btw for side questions you don't want added to the main conversation > subagents for research that doesn't need to stay in the main thread don't drag 40k tokens of debugging history into the next feature for absolutely no reason. 8. USE PLAN MODE + WORKTREES FOR BIG CHANGES for a big change, i'd use /plan before Claude starts editing. make it understand: > what has to change > what shouldn't change > dependencies > validation > where the risky parts are and if the work can actually be separated, parallelize it. /batch can split codebase-wide work into independent pieces and run them in isolated git worktrees. that is WAY better than having 5 agents editing the same branch and praying. 9. MAKE REVIEW A SEPARATE PHASE before i trust a meaningful change: > /diff > /simplify > fresh Opus 5.5 review > /security-review when it matters > tests + typecheck /simplify is especially useful because Claude can have multiple reviewers look for unnecessary complexity, reuse + efficiency problems after the code already works. the model that built the feature should not get the final vote on whether the feature is good. 10. BENCHMARK YOUR OWN SETUP save 5-10 tasks from your repo. that typically consist of: > tricky bugs > features with existing patterns > refactors without changing behavior > API changes > regression tests > large multi-file changes save the task + the command that proves whether it worked. then when you change your CLAUDE.md, skills, hooks, model routing or agents... run the same tasks again. otherwise you're just guessing whether your “better Claude Code setup” is actually better. models have gotten so good that at this point i'm way more concerned about the harness surrounding them.

AI labs have completely lost it this month... we're only 17 days into October and we've already gotten: > Claude Haiku 5.5 > Gemini 4 Argon > Muse Wonder + Wonder Full > Muse Spark Open > Mistral Large 4 > DeepSeek V4.1 Flash 1017 > Beam 1.0 > Nemotron 3.5 Super VL > Nemotron 3.5 Bolt > Ultramarine 1 Flash > LFM2.5 > Mellum 2.1 Thinking > Jais 3 > d1-ultra > Grok Voice 3.0 > GPT-Live 1.1 and we're STILL waiting on: > Fable 5.5 > Kimi K3.1 > Qwen 4 > DeepSeek V4.1 Pro > Mistral Large 4 weights > Nemotron 3.5 Bolt E4B and Astra 6.1 was literally supposed to be part of this insanity too before OpenAI held it back. AI tech is moving so freaking fast lol