
https://xitter.gq/t.co/TPhNX2FZ6X @01dotinc
Mars
Joined March 2021
- Tweets2.7K
- Following2.4K
- Followers316
- Likes32K
Alf Viktor Williamsen retweeted
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
Excited to announce that all super intelligence organizations have now jointly agreed to the ultimate in AI/SI safety:
Moving all testing to Delta Airlines flights, where accessing the Internet is utterly impossible!
Alf Viktor Williamsen retweeted
We built high-throughput materials labs in Menlo Park to create a loop between experiments and models. The labs generate fresh data, the models learn from it, and then help us decide what to try next.
Using only 1,300 H200s, plus months of our experimental data, we mid-trained and RL’d an open-source model to surpass GPT-6 Astra on our analysis benchmark. We call it Neon.
This is real footage from our lab. We’re focusing first on hard problems in materials science, including superconductors, magnets, and semiconductor materials.
Read our blog posts below.
Alf Viktor Williamsen retweeted
"Everything that needs to be said has already been said. But since no one was listening, everything must be said again." - André Gide
Alf Viktor Williamsen retweeted
People killed by AI: zero
People killed by governments: millions
Let's give control of AI to the governments and call it safety
Alf Viktor Williamsen retweeted
We’re introducing Kylace, for teams whose AI systems cost too much, fall short or need to scale.
We bring specialist expertise and take on the practical work of testing improvements against your requirements.
This builds on the research and production evaluations we’ve been publishing. Kylace makes that work available to your team.
kylace.com/
Deepseek just dropped v4.1 flash, fully open weights
it beats gpt 5.6 sol, opus 5 and every chinese model on coding and cybersecurity at ~86x cheaper cost per million tokens running at 420-507 tok/s = faster than gemini 3.8 flash
"smallest model in our new architecture family" btw
Alf Viktor Williamsen retweeted
I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives. More thoughts below.
Alf Viktor Williamsen retweeted
Replying to @ChaseLochmiller @OpenAI
GPT-6 Astra, trained on ~100K+ NVIDIA Grace Blackwell NVLink72. From ChatGPT to o1 to Astra in 4 years.
AGI has arrived. Congratulations @OpenAI team.
400K GPUs coming online next.
Alf Viktor Williamsen retweeted
Grok 5 is still deep in training on the massive Colossus II cluster.
Roughly 220k high-end NVIDIA GPUs are already online, with more being added.
Power draw is sitting around 1 GW and climbing.
The model is targeting the 6 to 10 trillion parameter range, which puts it in a completely different league from anything currently public.
Training has been running since at least January.
Elon recently confirmed it should arrive before the end of 2026 and will be trained on the full 25-year SpaceX engineering data set.
That combination of extreme scale plus proprietary real-world engineering data is the real story here.
This is not another incremental Grok 4.x update.
It is the much larger architectural and compute jump happening in the background while the smaller releases keep rolling out.
When it lands it should give SpaceX a serious internal edge on engineering, simulation and design problems that current models still struggle with.
Cursor is already running on the same cluster, so the jump in capability should show up there too.
No firm date yet, but the ambition on both compute and data is very real.
Alf Viktor Williamsen retweeted
This is really accurate.
As much as we all want to drink the vibe coding kool-aid, the reality is that 90%+ of all code in a company is maintaining and migrating existing stuff that is complicated and messy. You can definitely vibe code the remaining 10% but unless there is a working example of how to reinvent the 90%, AI will not live up to its potential.
We started 8090 and built Software Factory to focus on the 90%.
Many large enterprises now rely on us to help them migrate complex systems, rewrite old systems and maintain existing decisions.
Cheaper, faster and better.
Please consider trying it and seeing what it can do.
Every time I see a tweet saying “I can vibe code this in a weekend” - I think of the slack notification system..
It takes time, persistence and effort to get the details right.
Sure, a lot of simple workflows will get vibe coded away. And maybe you can put this in Claude Code and get the code right in one shot.
But quality, depth and great systems will still have value and take time. You can’t vibe code lessons.
Now and forever.
Alf Viktor Williamsen retweeted
Narrative violation: The largest AI data center uses roughly the same amount of water as two burger joints. There are over 200,000 fast food restaurants in the U.S. so this is a tiny amount.
Alf Viktor Williamsen retweeted
We’re launching Cocentives, a commission management system for insurance agencies.
Insurance agencies grind through commission statements from carriers every month. These statements arrive in different formats - Excel files, CSVs, PDFs - with different structures, different column names, different ways of identifying agents. Today, a leader or accountant at the agency has to turn all of this into usable numbers: how much each agent earned, from which carriers, for which policies.
This work takes hours per month at a mid-sized agency and does not scale very well. It’s manual, repetitive, and error-prone. It’s tiring work. Most agencies use spreadsheets. The person who maintains the spreadsheets becomes a single point of failure.
Cocentives replaces this process with an entirely new flow. As files arrive, forward them to Cocentives from your inbox, or upload them in the dashboard. From here, the files get analyzed and read deeply. Cocentives works through your files for you. It extracts the data cleanly, matches your agents across documents and history, calculates the totals, and produces the statements. All you have to do is review and send to your team.
Cocentives uses an advanced multi-analyst architecture to process and work with data. When you upload files, analyzing agents run in parallel to parse different formats simultaneously, extract structured data, and reconcile it against what the system already knows about your agency.
There are no templates to configure. Cocentives talks with you to infer structure from your data and documents themselves. When Co encounters anything ambiguous - an agent name it hasn’t seen before, a commission rate that doesn’t match historical patterns - it asks. Your answer teaches the system and cascades the changes for you. Next time, it won’t need to ask.
Each agency builds up what we call an organizational memory: the carriers you work with, your agent roster, your split structures, your exceptions and bonuses. After processing a few months of statements, the system knows your commission logic well enough to handle new files with less and less intervention.
You interact with Cocentives through conversation or the simple interface options. You can type natural language requests:
“Create a report with the files I sent you for February”
“Open last month’s report”
“Show me the breakdown for agent Mike Torres”
The system interprets these requests and makes action happen for you - opening panels, starting processing runs, surfacing specific data. This is not a chatbot bolted onto a traditional interface. The conversation is the primary way you navigate and control the application. From the ground up, Cocentives is built to be intelligent and fast. Instantly you will notice it is unlike any other software you have used for your agency.
There’s also a standard UI for reviewing results, approving statements, and sending reports to agents. Agents receive their statements via email with a link to an interactive portal where they can view their earnings and ask questions about specific line items.
Cocentives looks simple on the surface. But our simple interface hides a sophisticated and advanced system. We did this on purpose. We want to hide the complexity and only disturb you with the essentials. Cocentives does the rest.
If you want to maximize your agency performance, then Cocentives is perfect for you. Our software is improved every day to support your work. Go to cocentives . com for more.
In 2026, Claude Code could finally unleash the golden age of local and decentralized apps.
The reason is that Claude Code allows you to quickly clone any moderately complex cloud-based app into a decent local one that runs on only your files.
This is an important update to Obsidian founder Kepano’s concept of “file over app.” His argument was that files are portable (and hence more reliable) but apps are not.
The new information is that apps are suddenly portable too. That is: apps of moderate complexity without strong global network effects are suddenly easy to clone. The clone won’t be perfect right away, but it’ll be pretty good. And if the cloning dev sticks with it it’ll get better.
So: can we get a local open source Mac app for everything, operating only on your files? Maybe we can make that a reality.
File over app
File over app is a philosophy: if you want to create digital artifacts that last, they must be files you can control, in formats that are easy to retrieve and read. Use tools that give you this freedom.
File over app is an appeal to tool makers: accept that all software is ephemeral, and give people ownership over their data.
In the fullness of time, the files you create are more important than the tools you use to create them. Apps are ephemeral, but your files have a chance to last.
The pyramids of Egypt contain hieroglyphs that were chiseled in stone thousands of years ago. The ideas hieroglyphs convey are more important than the type of chisel that was used to carve them.
The world is filled with ideas from generations past, transmitted through many mediums, from clay tablets to manuscripts, paintings, sculptures, and tapestries. These artifacts are objects that you can touch, hold, own, store, preserve, and look at. To read something written on paper all you need is eyeballs.
Today, we are creating innumerable digital artifacts, but most of these artifacts are out of our control. They are stored on servers, in databases, gated behind an internet connection, and login to a cloud service. Even the files on your hard drive use proprietary formats that make them incompatible with older systems.
Paraphrasing something I wrote recently:
> If you want your writing to still be readable on a computer from the 2060s or 2160s, it’s important that your notes can be read on a computer from the 1960s.
You should want the files you create to be durable, not only for posterity, but also for your future self. You never know when you might want to go back to something you created years or decades ago. Don’t lock your data into a format you can’t retrieve.
These days I write using an app I help make called Obsidian (@obsdmd), but it’s a delusion to think it will last forever. The app will eventually become obsolete. It’s the plain text files I create that are designed to last. Who knows if anyone will want to read them besides me, but future me is enough of an audience to make it worthwhile.
Alf Viktor Williamsen retweeted
Effective prompt method for a new app with very high accuracy bar:
Simplified:
```
Message:
Requirements + Instructions + Role + What Not to Do + What to Do + Why + Where the Model Fits In The System
Pass entire schema with inline comments about what it all means. Three wide ranging examples that cover core edge cases and cases.
Pass the raw content coming from parsers.
Output:
Use the <thinking> tags to let the model breath before it outputs the real answer
Use the <json> tag to force model to output in perfect JSON
Force another <final_reflections_json> tag which forces model to add any note worthy things to the output which can be stored alongside the core JSON
```
See comment for open source link adding soon.































