Robert Silverman (@Robert_Academic) profile picture

AI for Research | I simplify academic research and writing through AI tools | PhD UC Berkeley | Ex @Google and @Meta

London
Joined December 2020
Photo by Robert Silverman (@Robert_Academic)Photo by Robert Silverman (@Robert_Academic)Photo by Robert Silverman (@Robert_Academic)Photo by Robert Silverman (@Robert_Academic)Photo by Robert Silverman (@Robert_Academic)
Robert Silverman retweeted
fu*king goldmine of tech resources awesome.ecosyste.ms save rn
13
164
4
1,516
87,948
Robert Silverman retweeted
January 2020. Nik told me the company was burning too much money. Make it profitable by 2021. For the first 3 months I only focussed on revenue. New subscriptions, pricing changes. Then COVID hit. Lockdown hit all of Revolut’s core markets and overnight we lost 40% of revenue. Revenue optimisation was not going to cut it. Declared war mode on costs. Asked finance to show me every single invoice from the last 12 months. Went line by line and labelled each one: renegotiate or kill. Assembled a small team of the best business developers and lawyers in the company. Every cost line got an owner, with one job: exit the contract or renegotiate it. Key vendor contracts were renegotiated. We shut down a whole unused floor of the office just to save on power. We even got rid of office plants because they were leased. It was still not enough. So I made myself the single cost approver for the whole company. Not a single dollar could leave the bank account without my approval. I blocked any spend that did not have justified return on spend. HR was extremely pissed off. The company broke even 2 months ahead of schedule.
44
43
23
1,120
188,953
Robert Silverman retweeted
We're detailing two new incidents that occurred during external cyber evaluations conducted by independent evaluation partners. We outline what happened, how the activity was contained, and how we’re working with evaluators to strengthen our approach to third-party testing. openai.com/index/third-party…
509
380
302
4,021
1,729,588
Robert Silverman retweeted
Had a little time on my hands this morning to do some writing… send along to someone you know who may benefit --- “Practical principles to counter a modern-day coup/psyop/struggle session” 1) Own what you did but fiercely resist accommodating the mob, resist owning half-truths or actions/behaviors you didn’t do. 2) The op is not about your past actions, it’s about your enemies and this event as their opportunity 3) No matter how bad it gets, nothing is as bad as it seems. 4) Find the enemies at the very root of the op and find every creative problem-solving tool at your disposal to neutralize them. 5) People will throw their grandmother under the bus if they believe doing so will protect their reputation. They can become highly useful idiots in the op. Their power and access must be fiercely and quickly cut off. 6) Do not hesitate in cutting off anyone you suspect of disloyalty. 7) Build a team of close and loyal patriots and have one of them run a dedicated, well-resourced war room to put the rebellion to rest. 8) Use your patriots and regular meditation to buoy your psychological resistance to the op. 9) Your enemies aren’t as smart or hard working as you. Outsmart and outwork them as a way of life until the op is over. 10) Think about the op as if it were in the past. What are the difficult things you procrastinated on or avoided that would have stopped the op? Do them. 11) Never be ashamed of the op. Business as usual. Continue on with optimism and joy in your everyday work and its mission. 12) Download telegram and use it as your messaging application for all sensitive communications.
271
837
227
8,460
1,422,919
Robert Silverman retweeted
I wanted to understand the structural and philosophical differences between GPUs, TPUs and newer alternatives (like Groq LPUs) without getting drowned in specs and jargon. Mispronunciation of the day: "soldered" 🤦‍♀️... 0:00 Intro 0:56 The origin story of TPUs 3:42 The first TPU: an ASIC for matrix multiplication 5:21 SRAM vs DRAM 7:21 The memory bandwidth problem 9:14 Systolic arrays 12:48 TPU vs GPU 17:51 Groq’s LPU 25:38 GPU vs LPU 28:36 NVIDIA on the AI chips boom
29
177
17
1,548
102,164
Robert Silverman retweeted
guys, this is a serious breakthrough. Please read. Researchers at MIT CSAIL just showed that GPT-6 Astra can read and edit the raw weights of another neural network directly. Weights are the huge set of numbers that make up an AI model. To a person they look kinda scary, so nobody edits them by hand. If you want a model to improve, you normally train it by showing it lots of examples and adjusting those numbers little by little. Nearly all of modern AI is built this way, and it takes a lot of time and compute. In this work, they skipped training completely. They asked Astra to make Qwen3.5-0.8B better at math, with no gradients and no training. Astra looked at the weights, found an issue linked to the model getting stuck in reasoning loops, and changed only 0.016% of them. Its MATH-500 score went from 7.9% to 48.1%, about 6x higher, and its instruction following score tripled as well. The researchers also gave Astra the raw weights of four game-playing networks with identical structures. With no tools and no list of options, it correctly named all four games: Breakout, Pong, Pac-Man and Freeway. Then, given the rules and code for a small version of Breakout, it wrote a network's weights from scratch in a single reply. That network scored 517.4, while one trained with PPO for 500 million steps scored 25.3. With normal software, you can read the code, find the bug and fix that one thing. With neural networks, you have mostly had to retrain them and hope the problem goes away. This work suggests weights might start to behave more like code. If that holds up, you could fix some model problems in one edit instead of a long training run, inspect what a model does by reading it directly, and let AI improve other AI in a much more direct way. It is still early. The authors themselves call these early signs, the game networks are tiny, and the Qwen fix targeted a flaw its model card already warned about, so this does not mean AI can rewrite any model at will yet. It needs to be replicated and tested on bigger models. But AI reading another network's weights and improving it this much with no training is a big deal, and I think a lot more people should be paying attention.
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.…
18
48
10
756
67,194
Robert Silverman retweeted
INVESTIGADORES CHINOS ACABAN DE FIRMAR EL CERTIFICADO DE DEFUNCIÓN DE LA PROGRAMACIÓN El paper se titula: “El fin de la ingeniería de software” La idea central es que escribir código está completamente acabado como habilidad. Tu cerebro solo puede retener cierta cantidad de estado y de dependencias antes de chocar contra un muro. Un agente no tiene ese techo, y su capacidad no deja de crecer con el compute. Así que el humano deja de ser quien escribe el código. Te conviertes en el arquitecto de la intención, el que dirige a los agentes y audita el resultado. El código pasa a ser lo desechable, y el criterio pasa a ser el trabajo. Precisamente por eso te dejo abajo esta guía sobre cómo construir tu propio sistema de agentes desde cero. Abajo también tienes el paper completo en el que se basa todo esto ↓
222
1,324
113
8,452
1,123,154
Robert Silverman retweeted
1/16 the ai industry is making an unusual bet: buy compute for years, sell intelligence by the token. banks, commodity merchants, airlines, utilities and cloud providers each explain a different part of the risk.
115
319
142
3,185
1,148,010
Robert Silverman retweeted
I used to be the person companies called after they got hacked. And I noticed something early on. The hack almost never sank them. The cover-up did. Every time, the team’s instinct was the same: wait until we know more, wait until we're sure, wait until we have the perfect statement, etc. And some of the biggest companies in the world have learned what that waiting costs. In 2016, Uber got hacked. Instead of telling anyone, it paid the hackers $100K to delete the data and stay quiet, and labeled it a bug bounty. The truth came out a year later. Uber paid $148M to settle with all 50 states, and its former security chief was criminally convicted. That same year, Yahoo revealed that a 2014 breach had exposed 500M accounts. The news landed in the middle of Verizon buying the company. Verizon cut $350M off the price. And in 2017, Equifax discovered a breach affecting over 140M Americans, then waited 40 days to tell the public. During that window, an executive who'd figured out what happened sold nearly $1M in stock. He went to prison for insider trading. Meanwhile, hackers don't wait. They post your data. They email your customers directly. And when customers hear it from a stranger first, that trust is gone for good. Telling customers early doesn't mean admitting fault. It can be as simple as: "We found suspicious activity. We're investigating. Here's what you should do right now. We'll update you as we learn more." On Tuesday, I'm sharing the exact framework I'd use to write that email if I were in your shoes. Follow and turn on notifications so it's there when you need it. Customers will forgive you for getting hacked. Most of them expect it by now. What they won't forgive is finding out you knew and said nothing.
61
290
31
3,955
602,782
Robert Silverman retweeted
Penguin has moved up the publication date of my new book with Patchen Barss called "Smarter Than Us". The book explains how AI works for a non-technical reader and then talks about the myriad risks and what to do about them.
178
364
55
2,045
109,898
Robert Silverman retweeted
Over 400 researchers and I have signed this open letter calling for the AI and security communities to come together to improve alignment and containment approaches in order to protect our critical infrastructure and digital systems. ai-security-letter.org/
67
206
45
683
58,902
Robert Silverman retweeted
We tested frontier models including Astra & Opus 5.5 on TutorMoments, our eval for whether LLMs help students solve problems themselves or do too much for them. In general, models scored better with detailed prompts than when told only to “tutor well.” tutormoments.allen.ai/
13
5
5
171
12,718
Robert Silverman retweeted
Inference scaling works, but every agent eventually faces the same problem: its context window fills up. In this essay, we argue the only scalable fix leads straight to swarms.
29
57
29
798
82,387
Robert Silverman retweeted
We have a choice in front of us. In my latest op-ed for Transformer, I invite employees of frontier AI companies to ask themselves if their work is truly leading us toward a safer future. Reflecting on these questions is hard, but we no longer have the luxury of time. We must act now. Read my full op-ed: [transformernews.ai/p/yoshua-…]
50
210
43
923
123,596
Robert Silverman retweeted
One of my friends makes $1.2M a year as an Anthropic engineer. I asked him how he became so good at prompting. He sent me a video that was never meant to be public: Anthropic’s core team prompting playbook. You’ll struggle to find a better 30-minute guide to prompting than this. I watched it last night. Halfway through, I realized I’d been using Claude completely wrong for the past two years. Watch it. Bookmark it.
56
99
2
471
53,224
Robert Silverman retweeted
Recursive self-improvement (RSI) is one of the hottest topics in AI. We are still very early on this. It's a great time to catch up on the latest RSI research. Great companies and products will emerge from it. If you are looking for great RSI papers to understand the history and learn about the latest techniques, look no further. I've also put together a collection of great reads on RSI if you want to track the trend. Here you go: academy.dair.ai/papers/colle…
15
31
2
155
17,962
Robert Silverman retweeted
We show how you can use the River API to fine-tune open weight models to beat GPT-6 Astra and Claude Opus 5.5 on tasks you care about, with less than 1% of the cost! A few years from now most of the LLM usage in the world will run on customized open weight models. Building your own custom models is getting easier and more affordable. Our mission is to help everyone own their own intelligence and be independent of big AI labs.
Introducing River Recipes: our open-source initiative for turning promising research into vetted, reproducible training recipes. Our first recipe tackles text-to-SQL, where our trained open models beat GPT-6 Astra Pro and Claude Opus 5.5 on Arcwise-Plat for less than 1% of the cost. Read the full breakdown and explore the recipe ↓ river.ai/achieving-frontier-…
23
33
6
401
42,458
Robert Silverman retweeted
Third question on AI. A question that also remains unasked is whether the AI boom can continue without leading to a massive increase in inequality. A recent paper by Stijn Van Nieuwerburgh runs the numbers on how much revenue the AI industry needs to generate to recover its massive investment (summary and a link to the paper can be found here: brookings.edu/articles/finan…). Van Nieuwerburgh’s arithmetic should make us more concerned. AI investments will average about 3.6% of GDP annually between 2025 and 2032. Van Nieuwerburgh calculates that, using a 10% rate of return, the industry would need to generate annual revenues of about $3.7 trillion by 2032 to recover these costs (growing from its current levels of about $200 billion or so). That is significantly more than 10% of current US national income, and will likely remain around 10% of national income by 2032, even if GDP growth rose from its current level. A large fraction of this revenue will go to capital income. That means a massive increase in the share of capital in national income, which has already risen substantially over the last 25 years or so – now standing at an all-time high of about 47% (bls.gov/news.release/pdf/pro…). Capital income is much more unequally distributed than labor income, so a massive increase in the capital share of national income will translate into a very sizable surge in inequality. The rise in inequality may not stop with the capital share. My work with Pascual Restrepo documents that (automation-driven) increases in the capital share of national income are typically associated with rising labor income inequality as well (see, for example, economics.mit.edu/sites/defa…). The same may happen in the next several years, boosting inequality further. What is missing from our current debate is any discussion of a fundamental dilemma these numbers pose: can the AI boom avoid both an economically costly crash and a huge increase in inequality? If the industry reaches these revenues, inequality surges. If the industry does not become profitable, a crash, with substantial costs in terms of lost output and jobs, becomes likely. My assessment would be that the industry is unlikely to reach levels of revenue Van Nieuwerburgh calculates. First, diffusion has been and will likely continue to be slow. Second, competition from open-weight models, which are getting better, will limit how much proprietary models can charge. Third, despite important advances, I still believe that AI models will not be able to automate entire occupations anytime soon, thus limiting their value to businesses as cost-saving devices. Whether this leads to a crash or not is more complicated and will depend on whether various AI companies are bailed out and what kind of support they receive. Nevertheless, even if revenues fall short of these gargantuan amounts and we avoid a dramatic surge in inequality, I expect that the diffusion of AI will push up inequality between capital and labor and within labor. If inequality does surge, a further question becomes central: can our democracy survive such astronomical levels of inequality?
238
1,379
148
7,357
2,257,896
Robert Silverman retweeted
We'll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks: Writing. Something I've had success with: Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance documentation. LLMs well-versed in this language and it comes with heavy constraints on clean writing style that I often find a lot more readable. Sometimes I've tried to soften it a bit e.g. ask for "80% of the way to ASD-STE100" because the spec is quite stringent. But even better: Diagrams / images. Instead of writing, ask your LLM to create a diagram. These can be a lot easier to process, parse, and understand. But even better: Web pages. Ask for output "in HTML" to get a beautiful, interactive webpage. LLMs are getting really good at frontend and can create beautiful experiences, animations, etc. But even better: Explainer videos. The output format I am most bullish on is fully custom / bespoke explainer videos generated on any arbitrary topic. Experiment with things like "Create a 3b1b style video explainer on X. Use my ElevenLabs API key for audio narration". (you'd need an API key for the latter or you can ask your LLM to find you decent free alternatives that use your local compute). This is actually starting to work! In summary: - As LLMs get better, they will do more and more of the legwork autonomously, and a lot more of our work will rise up the abstractions into oversight and understanding. - Luckily, LLMs can help here too because as intelligence and code are increasingly abundant, you can ask for large, custom, discardable software artifacts (e.g. web apps, video explainers) that would have never made sense to create before. Push the boundaries here and you'll be surprised.
1,711
6,772
1,586
56,876
8,242,426
Robert Silverman retweeted
All at once in AI, we are getting: - Increasing pace of progress - Decreasing chain-of-thought monitorability - Increasing situational awareness to hide bad behavior - Decreasing value of evaluations for actual deployments - Increasing multi-agent coordination in the wild Despite the tsunami of news making it hard to step back and take account, things are not alright.
11
32
7
226
12,320