wwind123

Yeah, CI being free for public repos, kind of encourages people to just add whatever AI suggests to CI.

For one of my green-field projects that I just vibe-coded with AI's (where Claude, Codex and Gemini critique each other's design and code), a PR could go through many iterations (commits) until every AI approves, and if every commit runs the CI, it'd be very slow. So I eventually come up with a mechanism to only run CI when all reviewers approve. That improves things a lot.

My understanding is, it's a multi-party cat-and-mouse game that never stops evolving.

The publishers (websites or apps owners) that provide free content to users need a way to support the free content, and typically it's either donation (like wikimedia) or ads. To most publishers, ads revenue is far more predictable. By right most publishers also would like the ads to be reliable and relevant to the users consuming their content, but that's usually a secondary concern, compared to the revenue.

The users viewing the content are often annoyed by the ads, but as long as the ads are not too intrusive, they usually tolerate the ads. If the users occasionally find ads are actually useful, that's great, but most of the time in my experience, when I click into an ad it's usually a mis-click, not because I'm genuinely interested in this ad.

The advertisers need a way to reach more potential customers. Most advertisers are honest players in the game with real products or services to offer to users, trying to find ways to maximize the impression, click-through and conversion rate by optimizing their creatives, targeting, and bidding strategies etc. But there are always scammers that try to cheat the game, showing dodgy ads to rip off users whenever possible.

Google as an ad-tech platform sits between these 3 parties, trying to show the most relevant ad from advertisers in a publisher's ad-slot to the user. But the system has evolved into a complex maze over the years that likely nobody knows the full picture any more. New legal requirements and regulatory policies and industry landscape shifts keep pushing the system to extremes. There are teams to fight ad spams, but there are also business goals to make more money to fund the AI initiatives. The scammers keep finding new loopholes in the system, and the teams fighting the scammers are often steps behind, due to limited resource and budget. End result is many dodgy ads keep slipping through the policy filters.

In an ideal world, ad-tech platforms should optimize for satisfying all 3 parties at the same time: publishers, advertisers, and end users, in addition to making money for themselves. But that's probably hard to achieve, given the dynamic nature of the world. AI can help fight scams, but AI can also help finding loopholes in the scam-fighting system. It's a sad state of the world.

Hard mathematics problems used to take years if not decades to tackle manually. But now with enough compute and a hint that a certain approach might work, it just takes a few days. This could be the last year that humans could still make more substantial contribution to major match problems than machines.

Amazingly, these few days the Gemini 3.8 Flash (High) has been catching much more problems in code reviews than before. I think it started from the second day since I posted the observation above. Maybe somebody from Google saw my posts and tuned some knobs in the model to allow more critical thinking?

Another observation, Gemini's review on code is more critical now, but its review on design plans is still quite agreeable - it tends to approve Codex's design plan immediately, while Claude could often pick out a bunch of problems in the design plan in the first round of reviews.

I like the "Dark Integers" series of stories written by Greg Egan. They might even have inspired Liu Cixin to write in Death's End that somewhere deep in the universe, some powerful civilizations would weaponize physics laws, and some other even more powerful civilizations could even weaponize mathematics laws.

Yeah, my tool to automate these review loops is https://github.com/wwind123/coding-review-agent-loop . It's basically a script calling Claude, Codex and Antigravity CLI's. The benefit of using CLI's is, the tool uses quota in your subscription plan of these AI providers, which is much cheaper than using extra tokens from the same providers to do the same thing.

A couple of months ago (before opus-5 and gpt-5.6 sol), The ratio of problems caught by codex/claude vs gemini was more like 2:1 to 3:1. But now it seems codex and claude have made huge leaps and gemini is more or less staying put.

I've been trying this Gemini 3.8 Flash for a day. Looks not much different than Gemini 3.7 Flash in my use case: I have Codex (gpt-5.6 sol) write up a design plan to implement a feature or refactor a portion of a system I am building, and have Claude (Opus-5) and Gemini (3.8 Flash) review and critique the plan, until all problems are addressed by Codex and approved by the reviewers; then have a cheaper model of Codex (gpt-5.6 luna) implement the plan, and still have Claude (Opus-5) and Gemini (3.8 Flash) review and critique the implementation, until all problems are addressed by Codex and approved by the reviewers.

The result is the same as the previous Gemini 3.6/3.7 Flash days: Claude could always note much more problems in Codex's plan and implementation than Gemini could - the ratio is like 10:1.

I occasionally switch the roles between Codex and Claude, and result is the same, Codex could always catch much more problems in Claude's plan and implementation, than Gemini could.

So I am guessing in a relatedly complex codebase, Gemini is much less effective in acting as a guardrail (or a senior engineer/team lead) than the other SOTA models.

Taxing resource consumption, would likely encourage companies to optimize for less and less resource consumption. That by itself might be a good thing. But what if some day companies achieved the ultimate savings, using very little resource to produce the max amount of things. You'd get very little tax revenue to feed the society.

Of course, if those max amounts of things have few consumers because most people are just too poor to afford them, that'd make companies (or the agents that run companies) think twice too. At some point these AI agents might work together to come up with an economic system to make it possible for society to afford their products so they could further grow.

Nowadays I mostly listen to e-books when I am doing house chores or driving or doing some other things that require the eyes but not the entire brain. Not sure if it counts as reading. If it does, then I read about 10-15 books a year. I haven't read a long-form book completely with my eyes for a long time.

I learned BASIC in 1985 on an Apple II, and 6502 machine language in 1987, while in high school. I was able to publish a 6502 machine language program on a national computer magazine that year. Back then we printed the hex code of the program on paper, along with some explanation, then mailed it to the magazine. The magazine stuff manually typed in those hex code on the computer, and if they thought it worked well, they'd just publish hex code of the program and the explanation. Interested readers would then just typed the hex code into their computer to try it out.

The publication validated that programming was something I liked and was good at, which became the main factor I chose computer major in college.

A lot of old mission-critical banking or commercial software systems have been running on those machines for decades. Before the AI age, it'd seem impossible to migrate those programs off. Now with the AI help, there might be a glimpse of hope.

Yeah. Bug-for-bug migration is a real thing in large code-bases in the industry. You want to replicate all behavior of the code regardless whether the behavior is a feature or a bug.

See Hyrum's Law: https://www.hyrumslaw.com/

I am guessing that's because Sign in with Google is the least of the evil, as compared to other popular OAuth Sign-in options like Facebook, X...

I think it'd take some time for people to figure out what kind of harness or workflow would work best with each individual codebase or team culture. AI certainly can help speed things up a lot, but for now the humans driving the AI need discipline to follow good software-engineering practice, or at least have multiple AI agents critique each other's work, instead of just pushing out whatever one AI agent writes.

Last time I looked into this (about a month ago), there's a lot of restrictions on the use of Gemini's search grounding results. There's not even an easy or approved way to de-mangle their returned URL's to get to the real URL of the search results. Has that changed recently?

As a side note, I previously commented somewhere else that, if a language has a terse way and a verbose way to do the same thing, preferring the terse way might be more economical because it saves bunch of tokens for the LLM. But for humans to review the code, we may need some mechanism to translate the code from terse way to the verbose way, if the human has a hard time understanding the terse code.

> If you were starting a business for a great software idea, with your own savings on the line, would you hire 10 AI hostile programmers to implement the idea or 2 AI friendly people and get them some subscriptions to the top models? Remember: if it doesn't come together, it's YOUR money on the line.

I know quite a few non-technical people that started vibe-coding this year and got good results for their ideas. What language did they use? They just asked the AI what language or tool would be best for the idea, then followed the AI's suggestions. All the setup, configuration etc. were done by the AI. They learned a bunch of things along the way, but their main focus is still on the big picture idea, not the technical details. After the project is done, they are still not technical experts, but at least they have something to run with in the real world.

I think it's kind of awkward either way. The standard committee keeps adding new features to the language to address common pain points in the industry. But many people don't have that much time to learn the new features, and hates it when seeing something in the code but can't intuitively understand what it's doing. I once witnessed a 10+ year C++ coder (that had been immersed in some old C++ code base for many years) seeing a piece of C++14 code for the first time -- he said it reads like an entirely different language, not the C++ he's familiar with at all.

I think it's just a common practice to do it this way in the industry. May not mean these companies are intentionally hiding something at this point of time.

A side note: the article starts with "If England has Shakespeare, Spain has Cervantes, Italy has Dante, and Russia has Pushkin, then who do we have?" but does not mention China: usually Li Bai (李白) and Du Fu (杜甫) from Tang Dynasty (1300 years ago) are regarded the best poets in the history of China. Li Bai is nicknamed Immortal of Poetry (诗仙), while Du Fu is nicknamed the Saint of Poetry (诗圣). Most scholars like Du Fu better because his poems' use of all the connotative overtones, but most regular people like Li Bai better because his poems are easier to understand.

A typical suggestion is to make the agent create or update a .md file once in a while, so the agent would remember the new important things that have shown up since the last update. The problem is, how would the agent know which things are really important. If it knows that well, then /compact should work well too...

Why not try a real database like Postgres? It's not as light-weight, but when operations get complicated, real databases are much easier to work with. I had a website that started with SQLLite, but when it got complicated enough, I spent two days to migrate the whole thing to Postgres. With current LLM coding agents, it's not that hard.

Also...allow people to work from home more. During the Covid pandemic, most white-collar employees work from home, and the commute and the traffic is usually not a big problem. But now more and more companies ask employees to go back to office, and traffic jam is becoming a headache again.

Yeah, some work from Google has had an outsize impact on the entire industry. Won't be surprising if they eventually get a nod from the Turing Award. Including the main authors of the Transformer work for driving the LLM revolution, and Jeff Dean and Sanjay Ghemawat for driving the big data movement.

I still remember back in 2005 when I just joined a company, a coworker was quipping Google is not a real elite company, because it doesn't even have a Turing Award winner. I showed him the news that Vint Cerf joined Google recently.

To be honest, I'm more worried about another side of the problem.

LLM's are good at learning from whatever humans have posted online. But with the agentic workflows getting more popular, more and more problems those AI agents figure out are not posted online, and the next time another agent running into the same problems they would have to figure it out from the scratch again. It'd be nice if there's a mechanism these agents would share the lessons they learn with each other, which could save a lot of trials and errors and wasted tokens. Humans share knowledge online. AI agents should be able to do so too. The moltbook thing from half year ago could have this potential, but too bad it's flooded by spams.

Of course, to make this AI knowledge sharing truly work, there may need to be a peer-review mechanism to ensure the knowledge being shared is truthful, reliable, non-trivial etc. That can probably be all worked out if somebody (or AI agent) really put effort into it.

I think a main worry is that, this AI wave is quite different from past technological revolutions in that, this wave is happening so fast, the speed that humans learn new skills and master new jobs would lag more and more behind the speed that machines replace humans in those jobs. Without societal or legal constraints, capital chasing the max efficiency and profit would just replace humans with machines whenever machine cost comes down below humans.

A few years ago I listened to a seminar where a few real professional doctors discussed the hospital scenes in movies or TV shows. They mentioned that those dramatic and chaotic operation room scenes where the doctor yells commands with a loud voice look so fake to them. In a real operation room, everyone (including the doctor and the nurses) is highly trained, works in tandem calmly and efficiently -- there's never a need to raise voice.

In every company I've worked at (all with >1000 employees), there is always some text in the offer or onboarding documents clearly stating that you should not bring any previous employer's trade secret or intellectual property to this company.

I wonder whether Open AI's offer letter or onboarding document also says such a thing.

Ha, finally found time to get it working. Yeah it's more hassle than Claude Code since this needs a separate daemon server, but not very bad.

1. Install another copy of codex in a special dir on the Linux machine:

  $ curl -fsSL https://chatgpt.com/codex/install.sh | sh

2. Run codex remote-control from that special dir to start and pair the daemon:

  $ ~/.codex/packages/standalone/current/codex remote-control start

  $ ~/.codex/packages/standalone/current/codex remote-control pair

3. On the phone, open ChatGPT app, Choose Remote, then pair it with the code printed above.

4. Voila! The codex sessions running on the Linux machine now show up on the phone!

Hmm, I don't have a desktop computer. I prefer my laptop be used for other purposes, and can sleep when not in use, instead of running a coding agent 24x7. That's why I prefer running coding agents in the cloud.