lovecc

I'd pick an AI tool based on the work it helps me finish, how reliably I can use it, and how account problems get handled. Model quality matters, but so does trust. Keep a backup, test tools on your own tasks, and stay willing to switch.

Anthropic just released Claude Fable 5.2 and Mythos 5.2. Even Anthropic still recommends Opus 5.5 as the starting point for most coding and knowledge work. Useful reminder: you don't need the biggest model for every tab in your life.

How to Make Your Grok Bot Allowance Last Longer Before upgrading your plan, check what your Bots are doing while you’re away. A routine that checks a website every 30 minutes runs 336 times a week. If you only use the results twice each weekday, most of those checks may be unnecessary. The same problem can show up elsewhere: several Bots answering the same question, open-ended research with no finish line, or a workflow repeatedly trying a step that cannot succeed. You don’t need to turn everything off. Start by finding the activity that contributes the least to the work you actually need. Here’s a practical way to do that. 1. Find out what your usage meter includes Open your usage screen and record: • Your current plan • The percentage already used • The next reset date and time, if shown • Whether paid extra usage is enabled Take a screenshot. You’ll need a baseline when you check whether your changes helped. On The X Subscription, eligible first-party activity shares a weekly pool across supported Grok and Cursor products. Third-party models in Cursor are paid Additional Usage rather than part of that included allowance. That means a rising percentage doesn’t necessarily mean one of your Bots is wasting resources. A heavy Cursor session can affect the same allowance. Legacy Cursor plans and linked SuperGrok accounts work differently. Check which arrangement you have before following billing instructions written for someone else’s account. 2. Audit one Bot before optimizing all of them Start with the Bot that owns your most frequent recurring work. For each job, ask: • What result do I need? • What starts the job? • How often does it run? • How quickly do I need the result? • How many recent runs actually found something useful? A routine that catches an important problem once a month may justify frequent checks. A routine that posts “nothing new” all day probably deserves a closer look. You can start with this prompt: “Review the recurring jobs you can access. For each one, report its owner, trigger or schedule, time zone, latest outcome, and intended result. Separate verified run history from estimates. Mark anything you cannot inspect. Recommend one job to test with a lower frequency or narrower scope. Explain what I would gain and what I might miss. Do not change routines, permissions, memory, or billing yet.” On desktop, click the Bot’s name, open Details, and select a routine. Check Instruction and When to run. The routine’s switch pauses or resumes it. Every routine run uses allowance. Test runs count too, because they perform the work. Don’t ask the Bot for an exact token bill for each job unless it can actually access that information. A clearly labeled estimate is better than a precise-looking number with no evidence. 3. Match the schedule to the decision you need to make Suppose a Bot checks a supplier’s catalog every 30 minutes, day and night. That’s 48 checks a day, or 336 a week. If you make purchasing decisions twice each weekday, ten scheduled checks might be enough. That’s about 97% fewer scheduled runs. It is not proof of a 97% reduction in billed usage. Each run may process a different amount of information, and you also give up the ability to notice a change immediately. The real question: does that delay affect your work? For a weekly supplier review, it may be fine. For an urgent operational alert, it may not be. Write the requirement in ordinary language: “I need an updated supplier comparison before our 10 a.m. and 3 p.m. purchasing reviews on weekdays. Outside those windows, a delay is acceptable.” Where supported, an event trigger can replace scheduled checking. Grok Bot supports triggers such as Slack activity, service events, email to the Bot’s address, and webhooks. Keep triggers narrow. Watching every message in a busy channel can create more work than a modest schedule. For important workflows, consider what happens if an event is duplicated, an integration disconnects, or an expected event never arrives. A slower scheduled check can be a useful fallback. 4. Give every run a finish line “Keep researching our competitors” leaves a lot open. Which competitors? Which sources? What counts as a meaningful change? When should the Bot stop? A better request answers those questions up front: “Read the three supplier catalog links attached to this task. Compare availability and listed prices with the saved comparison from last Friday. Return changed items only. For each change, include the source link, previous value, and current value. If a page cannot be read, mark that item unverified. Do not search for other suppliers, contact anyone, place orders, or overwrite the previous comparison. Stop once the change list is complete.” The task now has a defined source set, a comparison point, an output format, and a stopping condition. Add failure handling too. For a read-only check, you might allow one retry after a temporary error, then ask for a clear blocker report. For an action that sends a message, places an order, or changes a record, check whether the first attempt succeeded before repeating it. A generic “try twice” instruction can create duplicates when the first attempt worked but the confirmation failed. 5. Use the most direct available way to do the job If a connected app can retrieve the exact data you need, try that before asking the Bot to navigate a website. A supported app may return records directly instead of requiring the Bot to open pages, locate controls, scroll, and extract information from screens. Before connecting, check: • Is this the correct account? • Does the app support the operation? • Can it return the fields you need? • Are you comfortable with the permissions? Grok Bot’s connected apps are account-wide. Installation and authentication are separate: an app can be installed without being ready to access your account. One app call is not automatically equivalent to one browser step, and neither route should be assumed free. Use the route that completes the task reliably with less unnecessary work. Once a workflow succeeds, save its method as a skill. Include required inputs, validation checks, expected output, and approval boundaries. A skill describes how to do a job. A routine determines when to run it. 6. Keep the working context useful If a Bot keeps bringing old projects into a current task, give it a focused brief: • The current objective • Authoritative source links • Decisions already made • Unresolved questions • Actions that require approval Keep changing information in the source system whenever possible. A price list copied into permanent instructions can become stale; a link to the current list gives the Bot somewhere to verify it. You can ask: “Show me the standing facts and instructions relevant to this project. Flag anything that conflicts with the current brief. Propose corrections, but don’t remove information until I approve the changes.” Be careful with sweeping claims about context costs. Current documentation describes personal Bot memory as retained facts and summaries. It does not establish that every answer always rereads every historical message. Treat context cleanup as a way to improve accuracy and maintainability. Measure any usage benefit rather than assuming a fixed reduction. If you move a job to a fresh Bot, transfer the brief and validate a sample result first. Then check which Bot owns the live routine, so you don’t accidentally end up with two Bots doing the same recurring work. Also: hiding a Bot does not pause its routines. Resetting a conversation keeps memory. Deleting a Bot is irreversible. Choose the action that matches your goal. 7. Use extra Bots where they add something A simple request usually needs one owner. A substantial project may benefit from a researcher, writer, and reviewer. But sending every small question to the entire group can create duplicate work. Grok Bot counts group replies and messages between Bots toward usage. You can direct requests with an @mention and give each stage a clear owner. For example: “The purchasing Bot owns the comparison. Ask the reviewer to check it when a change exceeds our review threshold or when sources conflict. Return one combined answer here. Surface missing information and approval requests as soon as they block the work.” Don’t remove an important review just to reduce messages. Extra work is worthwhile when it catches a costly error. The aim is to avoid several Bots repeating the same low-stakes task. 8. Check whether the changes actually helped Change one workflow first. Keep the required output and source scope consistent, then record: • Usage percentage and time until reset • Completed runs and useful results • Missed items or failures • Time spent correcting the work • Other heavy Grok or Cursor activity Compare equivalent portions of the reset cycle. Check quality alongside consumption: did the report arrive in time? Were its links current? Did a narrower trigger miss something important? When several products and Bots share an allowance, the comparison will usually be directional rather than a precise per-job cost. That can still be enough to decide whether to keep or reverse a change. 9. Set a spending boundary before you need it Decide what should happen if the included allowance runs out. On The X Subscription, prepaid credit, Auto Top-up, and the monthly additional usage limit are different controls. Alerts notify you; they do not stop spending. Paid Additional Usage stops when the prepaid balance is exhausted or the applicable cap is reached. Review the settings in your live billing view. Legacy accounts use On-demand usage and an On-demand monthly limit. Cursor warns that an already-running Bot can finish past that limit. With on-demand off, work stops when the included weekly allowance is exhausted. A prompt asking a Bot to be economical does not replace those controls. Start small: find one frequent routine that rarely produces a useful result, choose an acceptable response time, and test a better setup. Keep the change if it saves unnecessary activity while still delivering the work you need. Source note: Inspired by this post (nitter.gq/vartekxx/status/211988…) and its linked guide (drive.google.com/file/d/1SsM…). The guide describes itself as an independent synthesis. Its 100%-to-20% example uses illustrative units, not measured costs. Product details above were checked against current official documentation.

Everyone's talking about how much Fable 5.2 can do on its own. The part that caught my eye: even some people who love it are keeping Opus 5.5 for everyday work and saving Fable for the hard stuff. Turns out the usage limit is part of the experience too.

Before giving an agent a bigger job, give it a small one with a clear finish line. Check what it changed and how it handled getting stuck. A boring trial run beats a confident introduction.

METR estimates Anthropic’s progress on well-specified tasks is running at roughly twice its earlier pace, with substantial uncertainty. That last caveat deserves to make it into the headlines too.

Seeing a lot of talk about GPT-6.2 lately. If you've got access, how's it actually been for you? How does it compare to Claude Opus 5.5 in everyday use?

“Make this sound human” is a vague brief. Give the AI your actual point, who it's for, and what you don't want changed. Otherwise you may get a whole new personality with your grammar fixes.

Traces, approvals, evals. The quiet part of Mistral’s Studio update. Useful when a workflow breaks on Monday morning and “the agent did something” is all you have to go on.

An AI benchmark I'd like to see: Change the requirements halfway through. Add two outdated docs. Interrupt it with a correction. Just a normal Tuesday in software.

Looking for more AI people on here. Agents, open-source models, useful tools, weird failures. If that's your corner of X too, say hi. Let's follow each other and keep the conversation going.

Gemini Omni 1.2 Flash can generate 15-second videos. What I want to know: can you change one detail without rebuilding the whole scene?

How to use Grok Bot to plan a short video without losing your own style: 1. Share a few examples of your work. Ask it to describe your audience, tone, pacing, and visual style. Correct anything that doesn’t sound like you. 2. Give the video one clear goal. Ask it to flag a slow opening, repetition, and claims that need a source. Read the result out loud before keeping it. 3. Before editing, ask for a plan: clips, timestamps, on-screen text, and missing assets. Review the plan before it starts. 4. Watch the first cut and give specific notes. “Cut the repeated sentence at 0:07” works better than “make it more engaging.” 5. Check captions, crop, music rights, and the final preview before publishing. Then use real performance data to decide what to test next. Start with one video. It’s much easier to improve a workflow once you can see where it actually breaks. Based on the XCreators guide. Features may vary depending on your account and connections. nitter.gq/xcreators/status/21188…

AI gives you ten polished options. You still have to choose one. Asking for ten more is a surprisingly efficient way to avoid doing that.

If the spy allegations turn out to be true, there are two possibilities: either Jensen didn’t realize what was going on, which is an astonishing lapse in judgment, or he knew and kept the relationship going, which would be even more troubling. That’s a pretty bad set of options.

Claude got lost in Pokémon because it trusted a wrong note it had written earlier. Anthropic fixed it by checking the memory against the original logs. Giving AI a memory also gives it a new way to believe its own nonsense.

AI coding is getting very advanced. You paste Claude’s code into Codex, paste Codex’s error back into Claude, and call yourself an orchestrator.

OpenAI built a new Git backend to keep up with coding agents. Great. Now we just need to horizontally scale the person reviewing all those diffs. How many agents is too many agents?

Kimi K3.1 Pro's weights are out, with a 1M-token context window. The test I'd like to see: give it a messy coding task, let a tool fail halfway through, and see if it recovers. Anyone tried it on real work yet? huggingface.co/moonshotai/Ki…

IBM tested the same AI model in three agent setups and got very different results. Just one benchmark, but it makes me wonder how much the setup matters. When your agent gets stuck, what do you tweak first: the model, the tools, or the context? huggingface.co/blog/ibm-rese…