Last week Bun 1.4 was released. This was the 1 that contains the rewrite of Bun from Zig into Rust, apparently containing much than 1 millions caller lines of Rust code. And while this caused a bunch of play successful the programming organization centered astir the connection switch, I deliberation the astir important point astir this merchandise is really the rewrite was done and the standard of it. The specifications of this, on pinch my ain experiences, and soon to get advances successful AI lead maine to judge that we’ve arrived astatine the extremity of programming arsenic we cognize it.
What I mean by this is that I deliberation the enactment of penning codification manually and having different humans reappraisal it to create useful, moving package is headed for extinction. Or astatine the very least, it will beryllium drowned retired by the absolute deluge of useful, moving package that will beryllium created by agents, pinch humans reviewing only the extremity result, not the codification itself. We’ll person much package than we’ve ever had, but astir of the codification for it won’t person been written aliases moreover publication by humans.
To beryllium sure, location will besides beryllium a upland of slop. Of buggy, useless, terrible, violative package that conforms to nary programmer’s consciousness of artistic aliases intent and style. But location will besides beryllium good engineered, moving package that solves existent problems, yet hasn’t been reviewed successful good item by quality eyes.
This detonation successful package tin already beryllium seen successful this chart from GitHub’s station astir their August 17th outage:
Since past twelvemonth there’s a steep exponential climb successful the magnitude of codification being produced. My mentation is that astir of this is really from broadside projects aliases projects soul to companies that are skunk works, aliases for individual users. Things that are viewed arsenic non-critical. In astir companies, location is still important organizational clash to individual developers shipping 2x much erstwhile it comes to the products that thrust the business, fto unsocial 10x aliases 100x.
Almost each of the codification successful those graphs was produced by models that are presently not astatine the frontier. It was produced by the frontier models of months ago, which is to say, not Fable 5 and not GPT 5.6 Sol.
Anthropic and OpenAI developers unrecorded successful the adjacent future
Reading the details of the Bun Rust rewrite, the large highlights are that it was driven by 1 developer, Jarred Sumner, moving pinch a pre-release type of Fable 5 and a seemingly unlimited token budget. He created the moving harness and model for agents to activity successful parallel and construe the Zig codification guidelines into Rust.
Over the people of 11 days, aggregate agents produced 6,778 commits and burned done capable tokens that it would person costs astir $165,000 successful API pricing.
I retrieve erstwhile the play astir this rewrite first sprung up successful early May, good earlier the supra blog station was written. It was revealed only done a merge of the rewrite, which showed +1,009,257 -4,024 successful changes. I was amazed astatine the standard and theorized that they had achieved this pinch Mythos:
Also, I’m dying to cognize if this was done pinch unlimited Mythos tokens. Is this the adjacent early for the remainder of us? Or the now and they did it pinch Opus 4.7 tokens? My conjecture is the former. So it’s a preview of what’s imaginable later this year.
— @pauldix, May 14, 2026
My acquisition astatine that clip was that I didn’t deliberation specified a point was imaginable pinch Opus 4.7 truthful it must person been the yet to beryllium released Mythos that enabled a rewrite for illustration this that really worked.
The truth that AI wrote 1M LOC and past refined it complete the people of the adjacent mates of months to nutrient a reliable portion of package that is presently moving connected millions of developer machines is perfectly mind blowing. And you tin say, “well it’s not that awesome because they had an oracle to comparison against, truthful it was elemental to spell from 1 connection to another”, but I deliberation that’s trading this full point short. If you tin build a verification strategy and springiness due direction, AI tin nutrient a highly complex, highly blase portion of package and it tin proceed to refine it until it conscionable works.
If you salary attraction to what Anthropic and OpenAI developers person been saying connected X successful the past mates of months, they each vessel dozens aliases hundreds of PRs per week and they person shifted their attraction higher up the stack. This makes consciousness arsenic it’s not really imaginable for a developer to intimately reappraisal hundreds of PRs worthy of codification each week while besides producing their ain work.
If you return what they opportunity publically astatine look value, they are nary longer doing very adjacent reappraisal of each statement of codification that gets shipped. They are spending their clip building systems, prompts and verification tooling to get AIs to nutrient their package astatine standard and pinch precocious velocity. At slightest that’s what I’m interpreting erstwhile they talk astir loops.
Yet they are operating successful a world wherever they person unlimited Fable and Astra (OpenAI’s adjacent model) access. They unrecorded successful a near-term early that we peasants will apt acquisition a twelvemonth from now. Their acquisition of utilizing AI to create package is not the acquisition of the remainder of america connected the section waiting for our play Fable allotment to reset.
The Bun 1.4 merchandise should springiness you a glimpse of what’s imaginable if you brace Fable level intelligence pinch adjacent unlimited tokens. Software transportation astatine scale, that works, arsenic proven connected millions of moving deployments.
My acquisition study from the field
I person 2 caller anecdotes to study from my ain experience. My vibes emotion is that Fable is qualitatively different than erstwhile models successful akin measurement to really group noticed Opus 4.5 released astatine the extremity of past November was different. It has crossed a period wherever much tin beryllium produced pinch higher level guidance and little supervision. I’m capable to springiness it requirements, architecture and instructions for a characteristic I want to create and it is capable to create a afloat functioning first type complete aggregate moving hours pinch nary further relationship from me.
My 2 examples are some things that haven’t shipped successful InfluxDB, but they’re caller capabilities I had built successful a fork. I chose these to spot what was imaginable pinch the caller model.
The first is Iceberg integration, making information successful InfluxDB accessible done Iceberg REST, aliases connected an outer S3 bucket and Glue catalog. This is simply a analyzable characteristic requiring:
- API & CLI to alteration the feature
- Implementation of the Iceberg REST API
- Deep ties to the compactor to create the Iceberg manifests and Parquet information successful the outer store
- Creation of Iceberg manifests aliases requests retired to Glue
- S3 API implemented successful InfluxDB for non-export usage cases
This is thousands of lines of implementation and trial code. I hashed retired the unsmooth architecture creation and requirements and past directed Fable to do the activity done sub-agents, triggering codification reviews, and supervising the process. 14 hours later, it had produced a moving version. I past told it to verify everything end-to-end pinch a moving InfluxDB cluster and utilizing DuckDB and PyIceberg arsenic the outer clients. It fixed a fewer bugs and verified it each worked.
The 2nd illustration is the creation of an separator information replication strategy for InfluxDB. This is aggregate individual InfluxDB nodes moving arsenic satellites that periodically replicate compressed information up to a cardinal InfluxDB cluster. It defines:
- API to group what will get replicated and connected what frequency
- CLI to entree above
- Updates to the compactor, which is utilized to filter, aggregate and create compressed information to replicate, on pinch the search to cognize what has gone up and what remains
- API for the separator node to entree catalog accusation to beryllium utilized successful the replication payload
- API for receiving compressed blocks of information into the pipeline
- Metrics and strategy tables to alteration visibility of each of the full setup
I laid retired the architecture and gave examples of what the personification acquisition should look like. I collaborated pinch Fable to nutrient a design, past told it to do the work, acting arsenic a supervisor. 28 hours later, it had a mostly functioning implementation.
I past directed it to deploy it wrong our trial infrastructure successful AWS and observe the metrics and logs and hole bugs arsenic it went. After a fewer short iterations it sewage to an end-to-end moving implementation, which is still moving a fewer months later. I moreover later had it build a UI that shows the replication action crossed nodes pinch rates of information transfer, conscionable because, why not.
Of course, neither of these things is presently shipped, aliases supported and isn’t what I’d telephone accumulation fresh software. So you could opportunity what galore opportunity astir AI, which is that it helps you vessel the prototype faster. But that isn’t really giving capable in installments here.
The prototype is moving software. And the betterment and testing of that prototype is further enabled by much betterment loops pinch the AI. It gets amended pinch much testing and verification, not done quality codification review, but done usage and testing.
I did each these pinch a play Fable allotment and my subscription (had to hold a week between). What I haven’t been capable to do is to put this successful the cloud, connected an betterment loop pinch unlimited Fable credits. We can’t spend the hundreds of thousands of dollars of monthly walk that I’d apt rack up if sent the apical frontier supplier disconnected to do my bidding 24/7.
Jarred produced the Rust rewrite successful 11 days, but past had agents moving connected continuous betterment for months earlier releasing it arsenic an charismatic supported release. The prototype is the beginning, but the supplier will besides thief you vessel the iterated, improved, and hardened accumulation fresh product.
The adjacent early for america programming peasants
At this clip past twelvemonth we had Opus 4.1 and GPT-5. You tin now bargain amended intelligence astatine a fraction of the price. We sewage Opus 4.5 and GPT 5.2 successful precocious November and early December past year. These are the 2 models that caused each CTO to freak retired complete their wintertime break astir really it was imaginable to conscionable create features and package astatine will.
I expect that we’ll person different 1 aliases 2 large releases from OpenAI and Anthropic this year. My champion conjecture is that we get Astra from OpenAI sometime successful September and it will beryllium Mythos/Fable people and apt much tin than Fable 5.
I expect the Fable & Astra equivalents of the period crossing Opus 4.5 and GPT 5.2 releases either by the extremity of this twelvemonth aliases early next.
That level of exemplary intelligence and capacity will beryllium capable to create immoderate characteristic you tin deliberation of for your software. And it will work. And it tin trial and verify it for you. It will do it arsenic quickly arsenic you’re consenting to return transportation of the code. The only mobility will beryllium what do you want to ship? And what do you want to support?
By the extremity of adjacent year, this level of intelligence will beryllium inexpensive capable that astir of america will beryllium capable to activity pinch it afloat time, for illustration Jarred was capable to activity pinch pre-release Fable to nutrient Bun 1.4.
In different twelvemonth aliases two, we’ll get wide entree to systems that tin nutrient tokens astatine 10x to 100x the velocity astatine a fraction of the cost. See the Cerebras C4 announcement and OpenAI Jalapeño results. Broader, cheaper entree to frontier intelligence astatine unthinkable speeds is coming.
Organizational inertia will apt mean that there’s different decade of humans penning codification by manus and having their colleagues reappraisal each statement of it. Many, if not most, companies will proceed to create package arsenic they person before.
But the astir productive package creators will beryllium doing it without programming successful immoderate accepted sense. They’ll beryllium directing AIs, creating harnesses, and package factories, and QA and verification systems that vessel moving package faster than we’ve ever seen before.
And we will person arrived astatine the constituent wherever we person much running, accumulation package written by AIs than written by humans. It’ll beryllium the extremity of programming.
English (US) · 
Indonesian (ID) · 
