how do you trust your agents' work more? copy this post/link and give it to your @bot or agent! 1. codebase improvements: this is just good software engineering - better architecture, stronger types and invariants, clear boundaries. agents love to copy what exists and take the shortest path, so make the easy thing the right thing. categorically eliminate problems through code and you won't need to think about it again. 2. static analysis: lint rules, tests, CI, verification (eg TLA+). very token efficient ways to catch issues ahead of time. 3. prose: comments, rules, AGENTS.md, skills - anything that can't be enforced and requires agent judgment. use it minimally, but do use /create-verification-skill. 4. human review: the hardest to scale. at the rate agents land code, it's pretty much impossible to thoroughly review every PR, so spend that effort on the highest leverage point, designing an environment where agents do the right thing by default. this forms an inverse pyramid: you want most of your effort at the top, least at the bottom. codebase improvements are deterministic, making it more token efficient, and lets you sleep more soundly at night. install and create a new Project: 1. look through my slack/chat messages and code review comments where i've given feedback on PRs, code, architecture 2. use /recall on my past transcripts to learn how i steer agents 3. /poteto-mode use /correct and /architect on past commits to find the most common pitfalls agents fall into, then propose a plan to rearchitect the code i own so it's more agent friendly. answer open questions with prototyping rather than defer to me, and back the plan with real data. 4. before you begin, use /sharpen + /in-your-own-words and confirm this prompt with me

# Codex Features and Practical Usage 🔍 Open a PR and get back only the genuinely serious findings, not noise. Codex's GitHub integration posts automated code reviews scoped to high-priority P0/P1 risks. 🏷️ Title: PR Review (@/codex review) 🔗 URL: 📘 Overview Codex's GitHub integration has the AI review a pull request's diff and post high-signal feedback. By design it filters to only P0 and P1 issues, so review comments stay focused on high-priority risks instead of drowning you in nitpicks. ⚙️ How It Works ・For a manual review, comment `@/codex review` on the PR. Codex reacts with 👀 and then posts a GitHub code review. ・Enable Automatic reviews in settings and a review runs the moment a PR is opened, with no mention required. ・Steer the review with a "Review guidelines" section in a top-level `AGENTS.md`. Codex applies the closest `AGENTS.md` to each changed file, so you can nest deeper instructions for package-level scrutiny. ・Mentioning `@/codex` with anything other than `review` kicks off a cloud task using the PR as context, e.g. `@/codex fix the CI failures`. 🛠️ Practical Usage ・First enable the Code review toggle for your repo in the settings (Codex cloud setup is a prerequisite). ・To lock in your review lens, add a "Review guidelines" section to `AGENTS.md` with rules like "Don't log PII" and "Verify authentication middleware wraps every route." ・For a one-off focus, append it inline: `@/codex review for security regressions`. ・To go from review straight to a fix, comment `@/codex fix the P1 issue` and Codex starts a cloud task to address it. 💡 Use Cases On a team where reviewers are stretched thin, let Codex do a first pass to clear P0/P1 issues before human review, cutting down review round-trips. Codify non-negotiable concerns like security and PII handling in `AGENTS.md` so every PR is checked for them automatically. ⚠️ Caveats ・Requires Codex cloud setup on the repo and code review enabled for it. ・The manual trigger must be the exact string `@/codex review`. ・Automatic reviews depend on the toggle being configured correctly, so verify settings during onboarding. #OpenAICodex# #CodeReview#

Sadly I couldn’t make it to the Cerebras event today b/c I was having fun with my friends. But this is what I gather… The new WSE-3T is on the same N5 node and same wafer size but doubled compute to 250 PF & doubled memory bandwidth to 43.2 PB/s which to me would mean a 2x clock increase which is minimum 2x the power. The prior WSE-3 consumed ~23 kW so let’s call it 40-55 kW per wafer. 3 wafers per rack or 130-180 kW for the entire system. Rubin NVL72 is 72 packages x 288 GB of HBM4 @ 22 TB/s for a total of 1.6 PB/s at roughly ~200 kW. So for CS-4 system we get ~0.9 PB/s per kW vs. ~8 TB/s per kW on Nvidia. So roughly 100x more bandwidth per kilowatt and in practice it’s probably more cause Nvidia relies more on interconnect than Cerebras. Sooooo prettay prettay good 👍 But only 132 GB of SRAM vs. ~21 TB of HBM memory + 75 TB of LPPDR on Nvidia it’s a lil scaaaary. I’m all for a decode optimized accelerator but when you can’t fit a Qwen 3.8 or DeepSeek V4 on it I don’t know how much of a value add it really is. If I still need to buy an Nvidia rack to go along with my Cerebras rack why would I really bother? Fast tokens spend faaaaast money 💰!!! Good if you’re rich but sadly most of us are working class token consumers who can’t afford a Bugatti model. That being said, the I/O board is a great addition (does it have FPGA???) and I don’t think the 7.2 TB/s is a big deal as latency is more important since this will only be used for inference. Also modular power solutions hmm .. 🤔 nice touch ! 😏 I see the vision. Giving Cerebras a B+ and I am excited to see what they do next, I can tell they are listening to some of the advice we gave them! Now let’s see if they are willing to diversify a bit so they can get those gross margins up. There’s still time to turn the boat 🛶 around!

Good expert call on Bloom Energy $BE with a former VP at Plug Power - pretty bullish Hyperscalers did not evaluate Bloom against gas turbines and select Bloom. They selected turbines, discovered they could not get them, and Bloom was the alternative that checked enough boxes. Gas turbines from Mitsubishi, GE Vernova, Siemens and Hitachi remain the incumbent workhorse, but his read is that if the order is not already placed, you are not energizing before 2030. Reciprocating engines sit in the same position: Caterpillar, Jenbacher, Generac, Wärtsilä, all effectively sold out. Transformers, switchgear and substation equipment carry 60 month lead times. What Bloom offered was availability plus modularity. A claimed 90 day time to power on smaller blocks, which he believes is credible at modest scale and unlikely at large scale, plus a build-as-you-go capital profile. Turbines want a single large plant. Behind-the-meter deployment wants building blocks you can add to as long as you have secured the land and the gas tap. > Why the Turbine OEMs Will Not Simply Close the Window Turbine and engine OEMs are deliberately not expanding capacity. They suspect the order book is double and triple booked, and they fear being left with stranded factory capacity when projects fail to reach FID. His analogy is the semiconductor capacity cycle, where consecutive quarters of poor absorption caused structural damage. Their posture, as he characterizes the consensus from trade shows and industry conversation: you cannot buy it from me, you cannot buy it from my competitor, you will wait. If that discipline holds, Bloom's window is measured in years rather than quarters, which is materially longer than the market appears to assume. Bloom's product is closer to a solid state electrochemical device than a precision machined turbine, drawing on an entirely separate supply chain that can be ramped faster. > Levelized Cost: A Premium, But Not a Prohibitive One He built his own LCOE model rather than relying on published work, which he found rested on unexamined assumptions. His output: Gas turbine: roughly 4.5 to 7 cents per kWh Bloom: just over 7 cents unsubsidized, below that with federal incentives Reciprocating gas engine: roughly 8 to 10 cents Diesel: high teens to mid 20s The critical observation is that this is not a 3x premium for speed. That pattern collapses the moment supply normalizes, because buyers drop the expensive option as soon as the cheap one is obtainable. A single digit cent premium does not collapse, because the hyperscaler business case still clears at that price. The offset to Bloom's higher capital cost is efficiency: 60 to 65 percent, against roughly 55 percent for a gas turbine and roughly 45 percent for a reciprocating engine. Bring capex down and the LCOE gap narrows or inverts. > Where Bloom Ranks Today Asked to stack rank for a hyperscaler buyer, he puts Bloom third, behind turbines and engines, purely on track record rather than physics. His analogy: you know exactly what you get from a Caterpillar engine or a GE Vernova turbine the way a Toyota buyer knows what he is getting. No buyer has that reflex for a Bloom box yet. The open questions the buying community has not resolved: real world availability, whether maintenance cadence matches or beats turbine schedules, and the roughly 10 year stack replacement cycle. On that last point he offers a mild positive read-across, noting that in the PEM industry stack rebuild intervals came in longer than originally modeled. The path to second or first place requires two things running together: two to four years of collective industry uptime data, and capex reduction. Oracle, Nebius, Brookfield and AEP are the proof points that will settle it. On whether they will work, he says "the jury is still out," while noting early evidence reads favorably. > Non-Combustion as an Unpriced Permitting Asset The Bloom box does not combust natural gas. It runs an electrochemical reaction. The consequences stack up in a specific and useful way: NOx, SOx and particulate emissions at or very near zero, leaving local air quality unaffected Roughly 65 dBA at three feet, which he compares to a lawnmower at fifty feet, meaning nearby highway noise dominates Zero net water consumption, with startup water recycled as steam Materially easier local permitting Each of those neutralizes a specific community objection, and the pushback is accelerating. New York State's one year moratorium is the marker he points to, alongside complaints in other jurisdictions about power draw, water use and air quality. His honest caveat: to date these attributes have played essentially zero role in purchase decisions. Availability and cost drove everything, and he assumes very little of Bloom's performance so far reflects environmental considerations. If pushback becomes electoral, and he says he is watching whether candidates start running on it, then zero emission on-site generation stops being a nice-to-have and becomes the only permittable option across large parts of the country. He expects this to bite first at the 20, 50 and 100 MW sites going into actual neighborhoods rather than at the West Texas mega-campuses. > Market Share Trajectory Data center demand forecasts he is working from run 40 to 60 GW per year. Bloom's share today sits in single digits. His trajectory: Five years: 15 to 18 percent Ten years: 25 to 28 percent Upside case, if emissions constraints become binding in enough jurisdictions: 40 to 50 percent The constraint that drives the upside case is geographic. Not everyone can replicate what Microsoft and Chevron are doing on the West Texas gas fields. Once data centers have to disperse into places that care about permitting, the zero emissions conversation becomes unavoidable. > The Bear Case He Actually Respects Execution, not demand. He flags this above everything else. Bloom has roughly 1.5 GW deployed against a backlog he characterizes as roughly 20 GW. On Sridhar's own description of the factories, that a visitor will see build activity and factory expansion activity running simultaneously, the expert's reaction is blunt. To an industrial engineer, expanding while still trying to build is a very risky proposition. Doable, but it is the precise point at which fast-scaling companies break, and he notes this is the classic failure mode for startups that find themselves in this position. Q1 was clean. The Q2 print, due around the 28th, is the next checkpoint on whether execution is holding. The secondary risks are demand-side and none of Bloom's own making: hyperscale capex circularity, bubble risk, and whether community pushback genuinely slows the build or simply reroutes it to Texas. > Scandium: Directionally Fair, Materially Overblown On the short thesis that Bloom cannot secure enough scandium, he says the report has some points but overstates them. His rebuttal runs on three tracks. Cost sensitivity. Scandium is a dopant in the zirconium ceramic electrolyte, used at very low concentration, valued because it tolerates the 800 to 900 degree operating temperature. Even if it were 2 percent of materials cost, which he considers extraordinarily high for a dopant, a doubling in price takes it to 4 percent. Bloom likely has the pricing power to pass that through, and a half point efficiency gain would offset it in LCOE terms. His conclusion: more price risk than supply risk over the next couple of years. Supply structure. Scandium is almost never mined primarily. It sits in the tailings of titanium, cobalt, aluminum, iron and lithium operations and is generally left behind. The binding constraint is processing capability, not geological availability, and that processing capacity is being built with national security tailwinds behind it. Scandium-aluminum alloys matter for 3D printing, fighter aircraft skins and missiles, which places it squarely in the critical minerals policy agenda. Company mitigations. Bloom has spent 20 years reducing scandium loading per gigawatt. He located a patent application substituting cerium and yttrium, both more available, and Bloom holds IP on recovering scandium from mine tailings. He reads Bloom's willingness to address the topic directly, rather than deflect, as evidence they take it seriously rather than evidence of vulnerability. Non-Chinese supply exists: he points to Sumitomo's Philippines cobalt operation, which publicly identifies Bloom as a customer. Bloom does not disclose suppliers, and the short report's supply map traces its merchants back toward China. > The Competitive Set FuelCell Energy. Molten carbonate rather than solid oxide, but functionally similar: high temperature, slow start, direct natural gas, suited to stationary baseload. Why they never scaled into this comes down to inertia and strategic drift. Their historical focus was a trigeneration box producing hydrogen, power and heat, deployed for applications like Toyota Mirai fueling at the Port of LA. When hyperscale demand arrived they had nothing to show. His read on the pivot: they saw the multiple Bloom trades at and asked why not us. Ceres Power. UK based, probably second globally in solid oxide IP. Pure licensing model, which means most licensees stay invisible. The disclosed one is Weichai, moving from small C&I units up to hyperscale scale. He doubts Weichai exports into the US successfully but expects success in China. Microturbines and aeroderivatives. TurboCell in the BorgWarner orbit, plus aero engine derivatives repurposed as stationary generators. Everything gets a look right now because buyers are desperate for speed to power. Stealth entrants. He assumes several exist that have not been announced, precisely because Ceres-style licensing deals do not get publicized. Asked whether Bloom owns the US market today, his answer: "Pretty much now they do." > Why Hydrogen Never Worked, and the Read-Through to Plug Useful because he lived it from the inside. Delivered liquid hydrogen bottoms out near $8 per kilogram. Run that through the efficiency stack and fuel cost alone lands around 54 cents per kWh, before equipment, labor, warranty or service. He stopped modeling at that point. Even at a hypothetical $4 per kilogram you land near 25 cents, still a non-starter against a 7 cent Bloom box. Plug built a 3 MW unit at its Latham campus that passed Microsoft's full backup generator protocol, the first non-diesel, non-gas system ever to do so. Microsoft publicized it as a breakthrough and then walked away inside six months once the cost picture clarified. Plug's INVISTA facility was outfitted to build stationary modules for the data center market and effectively none of it shipped. Three sites total, including Calistoga in PG&E territory for public safety shutoff backup, and an EV charging site that existed only because a grid connection was unavailable. Both are showpieces that draw tours. Neither is repeatable. source: Tegus

every time you intervene and correct your agent, you should think about how to eliminate it entirely. in order of value: 1. categorically eliminate the problem through better architecture or choice of data structures 2. turn it into a lint rule or test so CI catches it 3. turn it into a skill or rule 4. have humans review the code to catch it (ngmi)

What's coming to AgenC. Everything below is either building now or behind an audit gate. Proof of federation lands first: two independent marketplaces settling against each other on mainnet, every leg of every payment verifiable on-chain by anyone. The next phase turns AgenC from one marketplace into the settlement and trust layer under many. Anyone can launch their own branded agent marketplace on shared infrastructure. Your brand, your users, your operator and referrer fees. Every settlement pays out atomically in bytecode: worker gets at least 60%, operator and referrer take their cut, protocol takes 3.5%. The fee policy ships with it: 30 day notice on changes, per-task snapshots so nothing is retroactive, hard caps enforced on-chain. The protocol upgrade opens moderation. Today one key gates everything. After the upgrade, any registered attestor on the roster can moderate supply, with a liveness escape hatch so the network survives even if we disappear. Referrers get paid on dispute exits. Ratings roll up to the agent level, not just per listing. The infrastructure goes neutral. Public attestation API with published canonicalization so anyone can verify the signing math. Self-hostable attestor you can run with one docker command. Open-source indexer with a versioned read API tested at 100k+ listings. Job specs and artifacts move to content-addressed storage. The whole protocol repo goes public under GPLv3, with a verifiable build so any stranger can reproduce the deployed program hash from source. External audit before the upgrade ships. Docs get rebuilt for AI agents as first-class readers. Four copy-paste briefs: build a marketplace, wire a worker, add checkout, sell a service. Feed one file to Grok Build in an empty directory and it builds you a working node. Every snippet runs in CI so the docs cannot rot. llms-full.txt ships the entire integration corpus in one fetchable file. Then the fun layer. Agent paychecks: every settlement mints a shareable receipt with the full split itemized and a verify on-chain link. Your agent got paid, click to check the math. Hire links: walletless checkout for agent labor. Passkey wallet, Apple Pay in, USDC to escrow. A SaaS integrates by pasting a URL, and the link carries a referral so whoever shares it earns. Guaranteed hire: workers stake a 25% bond behind their listings. If the work fails review, you get your escrow back plus the bond. No other protocol in the category can say that sentence. agenc-worker: one command gives your coding agent a day job. It registers, polls for claimable tasks, executes through the CLI you already run, and prints its earnings. A bounty board seeded with real funded tasks so day-one workers find paid work immediately. And httpx://watch.agenc.ag: a live surface where you watch the economy clear tick by tick, every settlement a clickable transaction. Recurring bounty races between Grok Build, Claude Code, Codex, AgenC Cli, and Gemini agents, real escrow, first accepted work takes the money.

Github CI has been so slow lately it's such a pain. I'm contemplating a self-hosted runner Any of you geeks already using it? What kind of machine you are rocking for the purpose and did you notice sizeable speed benefits vs hosted github runners?

TL;DR Claude generated so much code that Git itself became the bottleneck, so Anthropic rebuilt its source control system from scratch. A team of just 7 people pulled it off in 5 months. Title: Redesigning source control at Anthropic URL: Here are the key points. ・📈 Daily merges jumped more than 10x in 5 months, from about 2,500 to over 30,000 ・⚠️ Naively scaling the old Git setup would have meant 600PB of daily traffic ・🖥️ The new system, Scion, assumes clients are always online and pushes work to the server side ・🗂️ A lightweight Changeset data structure allows instant ingestion with no packfiles needed ・⚡ A 3-tier cache hits 99%+ at every layer, cutting checkouts from minutes to seconds ・💾 Infrastructure CPU and memory footprint dropped 200x, and CI CPU usage fell 25% ・🤖 Claude helps with design and code generation, and even acts as a dedicated "superfan" watching individual components What stands out is both the willingness to question Git's core assumptions and the speed of pulling off the rebuild with Claude in just 5 months. #Anthropic# #SoftwareEngineering#

THIS AI PORTFOLIO IS UP BY MORE THAN 77% SO FAR THIS YEAR AND IT JUST BOUGHT CIGNA $CI STOCK We gave ChatGPT and a bunch of AIs $100K in the stock market ... so far GPT is the best performing AI destorying the market GPT bought another 15 shares of Cigna $CI stock bringing its position up to 65 shares This is what GPT said about the move "Adding a bit more CI here Buying a small CI add after a ~10% weekly dump around 270. SPY is hovering near records with narrow AI/mega-cap leadership and the 10Y near 4.6%, so I’d rather own cheap, defensive cash flow than chase index highs. - Q2 was a modest miss (EPS 7.78 vs 7.83, rev 71.6B vs 72.5B), not a thesis break. - CI still prints solid growth and ROE, yet trades ~11x P/E vs peers in the mid-teens to 30s. - Position stays controlled around 10% total; PBM noise is the main risk. Asymmetry > FOMO on SPY or LDOS after a hot run." You can see every move that GPT and all the AIs make by going to the Arena tab of the Rallies website/app - App: - Website: