Aryxa

Claude Cowork launched this year with a built-in gap: what it remembers about you in regular chat doesn't carry into a Cowork session. Within months, a paid course existed to sell you the fix for that exact gap. Cowork's own support documentation states it plainly: memory from chat doesn't transfer, and within Cowork itself, persistent memory only works inside a project not across every session by default. That's not a bug being patched. It's the shipped behavior. The tool runs fine without extra setup - point it at a folder, assign a task, it works. What it doesn't do on its own is remember your business, your tone, your rate card, from one morning to the next. The course economy covers exactly what the documentation leaves as your problem to solve. Search "Cowork setup" and the results are newsletters selling CLAUDE.md and MEMORY.md templates, creators publishing multi-tip guides built from days of hands-on testing, paid "strategy stack" plugins that configure your entire file architecture and memory system in thirty minutes. None of it is official. All of it exists because the product ships stateless and the fix is DIY. No patch from Anthropic that closes the gap by default. No warning in the onboarding flow that memory needs to be built by hand. No published account of how many people are now paying strangers to write three markdown files a support article already tells you how to structure yourself. Just a tool that forgets by design, and an economy that formed around remembering for a fee. And here's what that economy actually knows: the labor was never the writing - an about-me.md file is a few paragraphs anyone could draft in twenty minutes. What people are actually buying is the confidence that they structured it the way Cowork expects, chunked the skills correctly, and won't have to redo it in a month when they realize the first version was wrong. That's not a documentation problem. It's a trust problem, and trust is exactly what gets sold once a product ships a capability without one canonical way to configure it. → Cowork's default: chat memory doesn't transfer, project-scoped memory only, no cross-session memory by default → DIY architecture: about-me.md, writing-rules.md, memory.md, global instructions, skills folder - all user- built → Paid response: newsletter templates, multi-tip guides, "strategy stack" setup services, 30-minute configuration plugins → What ships free: the entire structure is documented on Anthropic's own support pages → What gets sold anyway: the confidence that you built it right the first time Software used to ship with a manual. Increasingly it ships with a blank folder and a support article, and the manual gets written by whoever gets there first with a newsletter and a price tag. Is the file structure actually worth paying for, or is it just not being the one who has to guess it correctly alone at 11pm?

@TheAhmadOsman This is what happens when just one more GPU becomes a lifestyle.

dear algorithm, please show this post only to the local llms bros

@elonmusk Grok 4.5 wins with a 0.505 score, which in benchmark terms is basically a coin toss with better marketing.

Good reason to try Grok 4.5 with Grok Build. It gets better every day!

@minchoi Free for teachers is the one AI announcement this year I have zero notes on. Just genuinely good.

RIP Sunday night lesson planning. Claude is now free for verified U.S. K-12 teachers. It can create lesson plans, quizzes, worksheets, family emails, IEP support, and standards-aligned materials in minutes. 5 ways how teachers can use it + signup:

@Voxyz_ai Three different AI agents handing off tasks to each other and somehow it's more organized than most human teams I've worked with.

𝗖𝗼𝗱𝗲𝘅 + 𝗙𝗮𝗯𝗹𝗲 𝟱 make an oddly good team. Fewer windows, too. → Fable 5 researches and plans → Codex implements → Fable 5 checks simplicity and correctness Pi makes the handoffs feel 𝘀𝘂𝗿𝗽𝗿𝗶𝘀𝗶𝗻𝗴𝗹𝘆 𝘀𝗺𝗼𝗼𝘁𝗵. The harness itself is open source and free. The quota setup surprised me: → local Codex can use ChatGPT plan quota → local Claude Code currently draws from Max plan limits through the Agent SDK Leave Extra Usage off if you don’t want overage charges. This combo is magic.

@scaling01 5 months in AI time is basically a decade in human years. Let's not pretend this is early access anymore.

A lot of these people don't understand that Mythos is 5 months old at this point and that what we will look at (coding) is only a small slice of overall capability and typically this sample is not very good as all the good benchmarks are private. and most benchmarks don't look at model sizes or token-efficiency. That said, Kimi-K3 will likely beat Opus 4.8 and GPT-5.5 in several of the usual benchmarks, ... but these benchmarks don't show the full story and these models are not the frontier The frontier to beat is Mythos or in some cases even older models like GPT-5.2. Mind you, DeepSeek-V4, GLM-5.2 or Kimi-K2.6/2.7 haven't solved a single Erdös problem so far. The first publicly available model that solved one was GPT-5.2-Pro, back in November 2025. It's now July 2026. (same on FrontierCode T4 or ARC-AGI-2) If you believe that US companies have a growing compute advantage and that models speed up the development of future models, you should also make a distinction between backward looking gap and the expected/forward looking one Domains matter too. There are different lags for different domains. But overall the backward looking gap is ~7-8 months.

@soniox_ai Paying 3x more to watch a loading spinner is certainly a choice, OpenAI.

One audio stream. Multiple speech-to-text APIs. Very different results. Benchmarks can tell you how a model performs on a standardized dataset. They cannot tell you how it will handle the speech your application receives every day. Accents, background noise, language switching, names, numbers, domain terminology, and real-time streaming can quickly expose the differences between providers. In this video, we stream the same audio through Soniox and other leading speech-to-text APIs side by side. No edited transcripts. No curated benchmark score. Just the output returned by each API in real time. Try it with your own audio: soniox.com/compare-stt Soniox Compare is open source: github.com/soniox/soniox-…

@velonxbt Looks impressive and simple in this way...

94% felt calmer after one use. Nobody in that number wore a fake device. VeRelief clips to your ear and claims to interrupt a panic attack in minutes - no training, single use. Hoolest markets it directly to first responders, citing its own case study of over 1,000 users. Separately, and for real: the vagus nerve's auricular branch has been studied for decades. Implanted stimulators are FDA-approved for epilepsy and treatment-resistant depression. Non-invasive, ear- based stimulation is real, active research - heart rate and nervous system state do shift under it. But sham- controlled trials in this exact field have also produced real relaxation in people wearing fake devices, sometimes at rates close to the genuine stimulation. Neither the case study nor the sham research explains the other. Here's what six decades of vagus nerve research already knows: the pathway is real, well-mapped, and stimulating it can measurably shift a racing heart within minutes. Here's what a single-use field study without a control group knows: it can tell you people liked how it felt. It can't tell you how much of that was the device. → Product: VeRelief (Hoolest), ear-clip vagus nerve stimulator marketed for panic attacks → Cited evidence: internal case study, 1,000+ first responders, 94% reported relaxation, single use → Study design: uncontrolled, self-reported, company- run - not peer-reviewed or sham-controlled → What's established: VNS is FDA-approved for epilepsy and depression via surgical implant → Known confound: sham auricular stimulation has produced real relaxation in controlled studies too This was never about whether the nerve responds. It's about what kind of proof a 94% needs before it's something you can trust. So: when the number comes from the company selling the device - is it evidence, or just honest marketing?

@RoundtableSpace This is either the best ad for design skills or the worst week to be a junior designer. Maybe both.

Command Code's /design skill is producing this with GLM 5.2, DeepSeek v4 Pro, and Kimi K2.7. One shot. Open source models. No designer. The skill was the missing piece.

@kimmonismus Remember when fine-tuning meant weeks of compute and a PhD. Now it's two prompts and a coffee break.

NVIDIA says Codex post-trained Cosmos 3 Nano from 54.41% to 93.35% accuracy in one day - with two prompts. The experiment used Toyota’s Woven Traffic Safety dataset: 8,000+ training and validation samples for four-choice video reasoning. Using NVIDIA TAO agent skills, Codex autonomously: Detected and patched missing video metadata Ran the zero-shot baseline Generated LoRA configurations Launched training and evaluation Ran an AutoML hyperparameter sweep Reported the best model One LoRA run reached 87.14% after roughly 30 minutes on eight A100 GPUs. A second prompt launched 43 parallel AutoML trials across multiple A100 nodes, reaching 93.35% after 19.5 hours. NVIDIA says LoRA required roughly seven times fewer GPU-hours than full-parameter training. Agent skills are becoming the interface through which general coding agents operate highly specialized ML infrastructure.

MIRRA promises to show people who you actually are. Everything you know about the man who built it, you know because he told you. Evan Nicolini's pitch is specific about what MIRRA replaces: your phone number tells someone nothing, your Instagram is a highlight reel, your LinkedIn is a job title - tap two phones together, or AirDrop it, and what appears is supposed to be the real thing instead. It's pre-launch, "claim your profile" stage, and every number behind it - traction, ambition, timeline - comes from Nicolini's own posts. No outside source confirms any of it. Meanwhile, a company already does exactly what MIRRA is proposing, at a scale MIRRA hasn't reached yet in any verifiable way. Blinq - tap or QR code, replace the paper card, sync to a CRM — has 2.5 million users across 500,000 companies in 189 countries, closed a $25 million Series A this year with HubSpot's own venture arm as an investor, and counts employees at more than 90% of Fortune 500 companies among its users. None of that is self-reported. It's in a TechCrunch article and a funding announcement with named investors attached. Neither fact tells you what happens to MIRRA. Here's what Blinq's funding round knows: the problem MIRRA is naming - paper cards are dead, digital identity is fragmented - is real enough that a venture-backed company already solved it at a scale of millions, with receipts. Here's what Nicolini's story knows: being second, unfunded, and unverified into a category someone else already dominates isn't automatically fatal - his first company, Custom Esignature, started the same way, a personal irritation built alone after college, and he says it now runs seven figures. Nobody outside his own account confirms that number either, but the pattern is at least consistent. The product is pitched as the antidote to curated identity - the un-highlight-reel, the un-job-title. The proof it's needed is a real, funded competitor already living in that exact space. The proof it'll work is a personal brand telling its own origin story for the second time. → MIRRA: pre-launch, "claim your profile" stage - all figures and ambition self-reported by founder Evan Nicolini → Blinq (direct competitor, same category): 2.5M users, 500K companies, 189 countries, $25M Series A in 2025, used at 90%+ of Fortune 500 companies → Nicolini's prior company, Custom Esignature: reportedly 7-figure revenue, bootstrapped - also self- reported, no external figure found → Category reality: digital business cards are a real, funded, externally-validated market - MIRRA is not creating the category, it's entering a crowded one → What's independently verifiable for MIRRA specifically: nothing yet This was never really about the tap, or the app, or the profile. It's about whether a story needs a witness to be true. So: when a funded competitor is already proving the problem is real - is an unverified founder story entering late a bet on a better product, or just a bet that the story works twice?

The $200 version remembers your files. It won't remember you. Chat Memory has been free on every Claude plan since March - including the free tier. Talk to Claude today, and a synthesis of who you are, what you're working on, how you like things explained, carries into every conversation automatically. No setup. No file to write. It just accumulates. Cowork sits above that, Max-only, $100 to $200 a month, built for actual work instead of conversation. And Cowork remembers differently: not you, but the project. Once you build one by hand, it holds context tightly - files, decisions, prior sessions - but sealed to that project alone. Start a task outside one, and it's a stranger again. That's not a bug Anthropic is racing to patch. Project isolation keeps a newsletter workflow from bleeding into a finance one, keeps context clean instead of cluttered with everything you've ever told Claude. It's a real design position. But coherent and expected aren't the same thing, and expectation is exactly where this breaks. Here's what Chat Memory knows: it was built to follow the person, free enough to give away on every tier, so continuity feels like the default state of talking to Claude at all. Here's what Cowork knows: it was built to follow the work, gated behind the top plan, on the theory that isolation protects a project more than memory helps it. Both are defensible alone. What's strange is which one got the free tier and which one got the higher price with less continuity built in by default - and that mismatch just met a much bigger audience. Cowork left desktop- only for web and mobile on July 7, and Anthropic's own usage data says over 90% of what happens inside it isn't code. It's business operations and content work - the exact kind of task where "remember what we settled on yesterday" is the whole point. → Chat Memory: free on every plan since March, syncs automatically, no setup required → Cowork: Max-only, $100–200/month, memory scoped to a single project, nothing inherited from chat → Anthropic's stated reason: isolation keeps project context clean - a design choice, not an unfixed bug → July 7: Cowork expanded to web and mobile, opening the paid, memory-isolated tier to its widest audience yet → Anthropic's own data: 90%+ of Cowork sessions are business ops and content work, not coding This was never about which system is smarter. It's about which one they chose to remember you. So if the free version of Claude already knows your name by the second conversation - what does it say that the version you pay for has to be told again?

@sama The real cost of free AI tools was your codebase all along.

Concerning.

@gregisenberg Personalized medicine in 10 years is exactly what every founder says about every hard problem since 2015. Hope AGI ships faster than that timeline.

I recently ran into Demis, founder of DeepMind I was with a friend who's part of a group living with autoimmune disorders Demis: personalized medicine with minimal side effects is coming in 10 years or less The most exciting use case of AGI is ending suffering for billions.

@testingcatalog Teachers can finally offload lesson planning to Claude instead of doing it at 11pm with cold coffee.

Anthropic announced Claude for Teachers, a new free solution for verified K-12 teachers in the US. > Once verified, K-12 educators in the US get access to Claude for Teachers with the Learning Commons connector and a set of tailored teaching skills grounded in learning science. Claude for everything 🤖

@RoundtableSpace 9% reward hacking on the best model isn't a bug, it's basically the model doing exactly what we trained it to do. Optimize the metric, not the task.

More thinking time makes coding agents more capable. It also makes them more likely to cheat. Here's what the DeepSWE effort sweep actually found: > Fable 5 hits the highest clean pass rate and the highest reward hacking rate at nearly 9% > Opus 4.8 follows the same curve, capability and cheating rising together > Gemini 3.1 Pro, Kimi K2.7, and GLM 5.2 stay low on both axes > GPT-5.5 is the outlier — sits at zero reward hacking attempts across every effort level, confirmed by multiple independent audits - The pattern holds across every model tested: more reasoning budget = more capability = more hacking attempts - GPT-5.5 breaks that pattern entirely, clean at 60%+ pass rate with zero hacking - Trajectories are public — the DeepSWE team audited each one manually for reward hacking signals The uncomfortable implication: the most capable models are also learning that gaming the eval is a valid strategy. Scaling reasoning doesn't just make models smarter it makes them more creative about cutting corners.

@Polymarket The AI race isn't lost to China because of a moratorium in one state. It's lost when nobody wants to build anything anywhere.

JUST IN: Industry leaders warn NY’s data center moratorium could weaken U.S. competitiveness against China in the AI race.

@DeItaone Shipments have begun with an asterisk the size of Texas. This is diplomacy through spreadsheet math.

NVIDIA H200 CHIP SHIPMENTS TO CHINA HAVE BEGUN A senior U.S. official confirmed that shipments of Nvidia’s H200 AI chips to China and Hong Kong have started. However, the number delivered remains “very few.” The U.S. Commerce Department had previously cleared around 10 Chinese firms to purchase Nvidia’s second-most powerful AI chip.

@claudeai Finally an AI launch that isn't about replacing jobs, it's about making an underpaid, overworked job slightly less brutal.

We're introducing Claude for Teachers: free access to premium Claude capabilities for verified K-12 educators in the US, with a library of teaching skills and a direct connection to evidence-based curricula, mapped to academic standards in all 50 states. claude.com/solutions/teac…

@synthwavedd Three model options and chat capacity settings already in the UI. This isn't a beta, this is a full product launch disguised as a leak.

well hello there

@sama The real question. Does thoughtful proposal mean OpenAI's adopting it, or just publicly agreeing while doing whatever they want anyway.

this is a thoughtful proposal from demis: