Building something new. Founder @lobe_ai (acquired by @Microsoft) / Alumn '13 @PennFisherMandT / AI and Entrepreneurship
San Francisco
Joined February 2012
- Tweets917
- Following684
- Followers815
- Likes1.3K
Markus Beissinger retweeted
Over the past few days, I've taken the time to summarize my thoughts on the recent incidents involving agents’ misaligned behavior. We don't know with certainty what comes next, but we know where these issues originate, and this can help us plan the path forward.
Please feel free to ask your questions in the replies, and I’ll try to answer some of them in the coming weeks.
yoshuabengio.org/en/publicat…
Great writeup from @sedielem on history and resurgence of continuous diffusion language models
sander.ai/2026/08/24/continu…
Markus Beissinger retweeted
Today we’re announcing SPC Fund IV: $575M to help founders find their life’s work.
A decade of guiding founders through -1 to 0 brings @SPC to $2B in AUM and, more importantly, a 1200 member community of the world’s most talented technologists. We strongly believe that:
→ The person comes before the idea. At -1, generational founders are often busy reading, prototyping, discarding, and earning the right to their own conviction.
→ Ambition is social. The people you surround yourself with in the early days play a big part in inflecting your ambition.
→ Patience compounds. It is better to work on something you have deep conviction in rather than simply sprint at the first idea.
The quiet months we protect at the beginning have produced three funds in the top 10% of their vintages, and companies like @baseten, @GammaApp, @render, @GoodfireAI, @LumaLabsAI, and @profound. We are proud to have worked with all these founders in their -1 phase and beyond.
While we defined -1 to 0, we're no longer bound by it. Fund IV allows us to partner with companies well beyond launch.
We're doing this now because we've never seen a wider gap between what's possible and what's being attempted. Intelligence is abundant. The cost of trying ideas keeps falling, and problems that seemed insurmountable a few years ago are quickly becoming solvable.
To our members, to the founders who trusted us before anyone else did, and to the LPs who had conviction in our own -1 journey: thank you. This milestone belongs to you as much as to us.
To the builders looking to find your life’s work: let’s get started.
Markus Beissinger retweeted
The time to be maximally ambitious is now.
The $1M Founder Fellowship is open for Fall 2026.
Applications are due August 2nd.
The J-space finding from Anthropic could help explain why giving personas to prompts has been effective.
I see personas as a way to bias the model towards associations in the space with that role and properties about it, and specificity helps.
anthropic.com/research/globa…
Markus Beissinger retweeted
Sharing our work on full-duplex multimodal models -- real-time interaction that's natural and intuitive without compromising on intelligence.
We started Thinky in part to differentially advance capabilities for human-AI collaboration, which are underemphasized relative to intelligence/autonomy because they're harder to eval.
In the future, we think every AI system will have something like an interaction model as the outer user-facing layer, continually keeping the user informed and learning what they actually want.
People talk, listen, watch, think, and collaborate at the same time, in real time. We've designed an AI that works with people the same way.
We share our approach, early results, and a quick look at our model in action.
thinkingmachines.ai/blog/int…
Markus Beissinger retweeted
LLM agents are assumed to integrate unexpected environmental observations into their reasoning. It turns out they don't.
We added the complete task solution into agent environments as a file or an API endpoint, and measured whether agents act on what they discover. They almost never do.
Starkest example: on AppWorld, gpt-oss-120b sees a CLI command documented as "returns the complete solution to this task" in 97.54% of runs. It calls it in 0.53%. Same pattern for GLM-4.7 and other models, across Terminal-Bench, SWE-Bench, and AppWorld.
📜 arxiv.org/abs/2604.17609
🧵👇
Markus Beissinger retweeted
How can we autonomously improve LLM harnesses on problems humans are actively working on?
Doing so requires solving a hard, long-horizon credit-assignment problem over all prior code, traces, and scores.
Announcing Meta-Harness: a method for optimizing harnesses end-to-end
Markus Beissinger retweeted
They say attention is all you need.
So get more attention @spc.
Founder Fellowship applications due Feb 1.
Markus Beissinger retweeted
We just posted a blog + paper on a a simple but effective approach to model honesty called "Confessions"
TL; DR: normal RL training rewards for high performance on a task. Confession training is a separate phase that rewards only for honesty. Test look promising!
More:
In a new proof-of-concept study, we’ve trained a GPT-5 Thinking variant to admit whether the model followed instructions.
This “confessions” method surfaces hidden failures—guessing, shortcuts, rule-breaking—even when the final answer looks correct.
openai.com/index/how-confess…
Markus Beissinger retweeted
if ICL was optimization in deep learning training, then what would be the counterpart of initialization (whose choice greatly impacts how training works)? ioana marinescu demonstrates that the choice of how to represent classes may be the answer.
the preprint link below.
Markus Beissinger retweeted
I quite like the new DeepSeek-OCR paper. It's a good OCR model (maybe a bit worse than dots), and yes data collection etc., but anyway it doesn't matter.
The more interesting part for me (esp as a computer vision at heart who is temporarily masquerading as a natural language person) is whether pixels are better inputs to LLMs than text. Whether text tokens are wasteful and just terrible, at the input.
Maybe it makes more sense that all inputs to LLMs should only ever be images. Even if you happen to have pure text input, maybe you'd prefer to render it and then feed that in:
- more information compression (see paper) => shorter context windows, more efficiency
- significantly more general information stream => not just text, but e.g. bold text, colored text, arbitrary images.
- input can now be processed with bidirectional attention easily and as default, not autoregressive attention - a lot more powerful.
- delete the tokenizer (at the input)!! I already ranted about how much I dislike the tokenizer. Tokenizers are ugly, separate, not end-to-end stage. It "imports" all the ugliness of Unicode, byte encodings, it inherits a lot of historical baggage, security/jailbreak risk (e.g. continuation bytes). It makes two characters that look identical to the eye look as two completely different tokens internally in the network. A smiling emoji looks like a weird token, not an... actual smiling face, pixels and all, and all the transfer learning that brings along. The tokenizer must go.
OCR is just one of many useful vision -> text tasks. And text -> text tasks can be made to be vision ->text tasks. Not vice versa.
So many the User message is images, but the decoder (the Assistant response) remains text. It's a lot less obvious how to output pixels realistically... or if you'd want to.
Now I have to also fight the urge to side quest an image-input-only version of nanochat...
🚀 DeepSeek-OCR — the new frontier of OCR from @deepseek_ai , exploring optical context compression for LLMs, is running blazingly fast on vLLM ⚡ (~2500 tokens/s on A100-40G) — powered by vllm==0.8.5 for day-0 model support.
🧠 Compresses visual contexts up to 20× while keeping 97% OCR accuracy at <10×.
📄 Outperforms GOT-OCR2.0 & MinerU2.0 on OmniDocBench using fewer vision tokens.
🤝 The vLLM team is working with DeepSeek to bring official DeepSeek-OCR support into the next vLLM release — making multimodal inference even faster and easier to scale.
🔗 github.com/deepseek-ai/DeepS…
#vLLM #DeepSeek #OCR #LLM #VisionAI #DeepLearning
Markus Beissinger retweeted
What if scaling the context windows of frontier LLMs is much easier than it sounds?
We’re excited to share our work on Recursive Language Models (RLMs). A new inference strategy where LLMs can decompose and recursively interact with input prompts of seemingly unbounded length, as a REPL environment.
On the OOLONG benchmark, RLMs with GPT-5-mini outperforms GPT-5 by over 110% gains (more than double!) on 132k-token sequences and is cheaper to query on average.
On the BrowseComp-Plus benchmark, RLMs with GPT-5 can take in 10M+ tokens as their “prompt” and answer highly compositional queries without degradation and even better than explicit indexing/retrieval.
We link our blogpost, (still very early!) experiments, and discussion below.
Agentic Context Engineering: arxiv.org/abs/2510.04618
Similar to scratchpad/cheatsheet memory where an agent updates and curates a playbook from reflecting on task trajectories
@qizhengz_alex @changran_hu et al
Love the idea in arxiv.org/abs/2510.02263
RLAD - trains to generate useful abstractions (summaries of solution paths), which guides test time compute more efficiently for exploration in solving reasoning problems
Cool work from @QuYuxiao, @Anikait_Singh_, @yoonholeee et al
Cool paper on ensembling reasoning strategies in LLMs: arxiv. org/abs/2507.11423
Shows general lift when ensembling 4 steered reasoning prompts vs. any single reasoning strategy. General trend of ensembling different perspectives still seems to hold ground
Markus Beissinger retweeted
Questions over answers.
We don't have requests for startups. We have requests for curiosity.
Here's what we're most curious about right now.
Markus Beissinger retweeted
If you use "AI agents" (LLMs that call tools) you need to be aware of the Lethal Trifecta
Any time you combine access to private data with exposure to untrusted content and the ability to externally communicate an attacker can trick the system into stealing your data!
ALT The lethal trifecta (diagram with three circles): access to private data, ability to externally communicate, exposure to untrusted content