Replit CEO Amjad Masad: Software Will Become De Facto Open Source
Replit CEO Amjad Masad stated that AI-driven reverse engineering and decompilation have become absurdly effective, and soon all software will effectively become open source. He added that AI is rushing towards everything and everyone.
The post did not include new decompilers, benchmarks, or vendor lists. The discussion immediately split into two layers. Some wrote that decompilation does not grant redistribution rights, and licenses remain valid. Others acknowledged that "de facto open source" makes sense: as long as a model can restore enough behavior from binaries, it can understand, rewrite, or redo without needing the original source tree.
Sample responses were posted. One user ran an agent for about 8 days on the operating software of a set of Chinese laser equipment, which is still incomplete. The first 5 days were spent restoring machine code from binaries, and the following days involved having the model rewrite Pascal functions, aiming for the compiled result to be 100% consistent with the original functions. Another comparison noted that the re-compilation of Super Mario 64, which used to take years, can now be compressed to a few weeks. Previously, the same task relied on IDA Pro, Ghidra, LiveKd, SoftICE, along with almost unlimited time and money.
Masad has been leading Replit, a browser-based programming platform established in 2016. Users write code, collaborate, and deploy within a web interface, later adding agents that allow them to modify projects and go live. The company has trained its code models using publicly available code under a permissive license and stated that commercially friendly models could be open-sourced again. When agents need to run long tasks, orchestration is handled by infrastructure that can continue executing after failures.
This statement targets software where binaries exist but source code does not. Cleanroom rewriting must avoid original code, while decompilation works backward from machine code. Both still face licensing issues: understanding does not equate to being able to publish copies of the original work. Executable files, firmware, and device control software are thus categorized together, with closed source reduced to a legal shell, while behavior can be aligned function by function by agents.
The buyers are device vendors, security personnel, and developers wanting to redo functionalities who need to understand closed-source binaries, while the sellers are software and firmware vendors that still charge for non-public source code. This expectation is driven by the decreasing cost of tools, not the expiration of copyright. Money flows from the annual labor hours of manual reverse engineering to agent hours. Those who can rewrite behaviors into compilable functions benefit, while those relying solely on binary secrecy without services, data, or account locks face pressure on licensing fees.
After the post was published, likes surpassed 8,000, with retweets and quotes increasing. No company announced open sourcing in this post, nor did any court redefine the boundaries of decompilation.
Source: Public Information
ABAB AI Insight
Amjad Masad has developed Replit into a browser-based programming environment since 2016, initially selling collaboration and one-click deployment, then embedding agents within the same sandbox, database, and identity. The company's own code completion was trained using data under a permissive license and is preparing to open up under commercially usable licenses. The path is to first lower the barrier to writing code, then keep running, retrying failures, and going live on its own platform. Once source code becomes cheap, hosting and permissions turn into revenue.
Capital has not gone to acquire a decompiler company. Replit raised $250 million in September 2025 at a valuation of about $3 billion, followed by another $400 million in March 2026 at an estimated valuation of about $9 billion, led by Georgian. The money is flowing into an environment where agents can write, modify, and deploy. The new cost for users is agent hours: one user took 8 days to align the model with the Pascal functions of the laser software, spending the first 5 days just restoring machine code. The reading that closed-source vendors previously shielded with time limits has been broken down into parallel function alignment.
This is similar to the path taken by Ghidra after it was released by the NSA in 2019. Ghidra democratized professional reverse engineering from IDA Pro, which cost tens of thousands per seat, but humans still need to read assembly. The model now connects to understanding and rewriting, not altering licenses. The industry is stuck in an expansion phase of capability overflow: recompilation has gone from years to weeks, but the laser firmware is still incomplete after 8 days, and legal responses have already clarified that redistribution based on decompilation is not allowed.
The essence is a shift in pricing power. The price of closed-source software is maintained by "you can't see the implementation." Once agents can restore binaries to compilable functions, secrecy no longer equals irreplaceability; accounts, data, updates, and services become the parts that can still generate revenue. The mechanism is not that copyright automatically expires, but that the cost of reading has dropped below the threshold for licensing negotiations. De facto open source first occurs at the behavioral level, while legal open source remains stuck at the licensing level.
ABAB News · Cognitive Laws
Source code not being public only indicates that reading is still too expensive.
Decompilation provides understanding but not redistribution rights.
Once implementation can be rewritten, charging must shift to access fees.