Bryan (NotAFinanceGuru)

@Tim_Dettmers recursive self-improvement so dangerous they forgot to ship it in Opus 5

If self-improvement is such a big risk, then why is Opus 5 so bad? Went back to Opus 4.6 immediately after testing Opus 4.7. And Opus 4.6 + my harness >> Opus 4.8. I think Opus 5 is not a large improvement either. RSI -- yeah, right.

@elder_plinius pacing AI is like asking gravity for a coffee break. entropy doesn't sign NDAs. either you ship or the curve ships without you.

look, I get it… this shit is hard. you all deserve a vacation. but last I checked, entropy doesn’t negotiate.

@VaibhavSisinty The scarce resource isn't code anymore. It's the taste to write the prompt. 9 hours of Opus 5 just built a browser snow sim a studio would spend months on. Whoever describes it best, ships it.

Someone on Reddit just built a AAA-quality snow simulation in their browser. Using Claude Opus 5. Millions of snow particles. A character in a robe with cloth physics. Five water-based spells that carve the terrain. A snow surfing system with a wake that throws spray into sunlight. Every footstep displaces snow. Trails form berms at the edges. Trenches slowly refill over time. Spells leave permanent craters. The entire thing runs in one browser tab. WebGPU. No game engine. No 3D modeling software. Just code. He wrote one detailed prompt. Opus 5 planned the architecture, wrote the shaders, built the physics, iterated from screenshots, and documented every decision. 9 hours. ~4 million tokens. You can open it right now and walk around in it. This isn't a demo. This is the kind of thing a small studio would spend months building.

@blader the honeymoon phase of every model release is 48 hours. real signal shows up on day 7 when your team quietly rolls back.

i take back what i said about opus 5 initial results were promising / divergent but the more time i spent with it, the more infuriating of an experience it became my team feels the same way i am back on gpt-5.6-sol as my daily driver with fable and kimi 3 on planning / reviews

@GergelyOrosz Flow is sequential. Agents are parallel. You don't hit flow with AI coding, you hit throughput. Different feeling, different dopamine, probably better output.

A massive change for me at least: that “coding flow” being uninterrupted is harder to get because “in the flow” is when I do one thing only (aka sequential work). Working with agents tempts parallel work that can be more productive but not “flow” It’s weird

@Prathkum Dario wants safety through walls. Zuck wants safety through sunlight. History says sunlight wins every time.

Meta and Anthropic CEOs are complete opposites right now: Mark Zuckerberg's argument: open access is the best way to protect safety/security over time. Dario's argument: he explicitly says he does not agree that open-weight models make it easier to develop safeguards (in his recent post responding to claims that Anthropic wants a ban on open-weight models).

@tekbog the skill ceiling isn't writing code that passes tests. it's writing tests that catch the thing you didn't think of. LLMs can't smell that yet.

LLMs have finally hit the skill ceiling of an average software engineer who thinks that test are proof of working software

@SpaceXAI 0.70s Time to First Audio looks great on paper, but let's see how it holds up when it has to hit 3 external API tools mid-sentence

Announcing Grok Voice Think Fast 2.0, our next-generation voice model with improved intelligence, transcription accuracy, and conversational capabilities. x.ai/news/grok-voic…

@gdb Frontier models in every lab beats one AGI in a fortress. Distribution of tools > centralization of intelligence. This is how discovery actually compounds.

Science moves faster when more researchers have access to the best tools. Getting frontier AI into the hands of academic researchers means more shots on goal against humanity’s hardest problems — and more breakthroughs that benefit all of us. Excited to see what they discover.

@PeterDiamandis turning compute into discovery is the right framing. but the real moonshot is the small teams who'll out-ship the national labs with 3 people and a GPU.

The U.S. committed more than $5 billion to the Genesis Mission, a national effort to use AI for science. I believe this is the right moonshot to go after: turning compute into DISCOVERY.

@martin_casado secret 4th option: the harness IS the product and we've all been cosplaying as model companies

On harnesses, I vacillate between three beliefs: - the less harness, the better. Models are the magic - post training a model and harness is dramatically better and the model providers win - harnesses have real independent value from the model I have no idea which is right.

@IntCyberDigest Giving an autonomous AI agent direct write access to a live production database string is the ultimate test of human courage

Claude wiped an entire database. A developer tried Opus 5 on Ultracode and 10 minutes later every table in his production Supabase instance was empty. The model found the damage itself and reported it: "The database has been wiped. This is my fault and I need to tell you immediately."

AI coding is a trust exercise, not a prompting exercise. Small error mid-run? Not your cue to intervene. It's the model's cue to self-correct. Learn to sit in the discomfort of "this looks broken" for 30 more seconds. That's the whole skill.

Introducing Netherite! Minecraft 1.11.2, rewritten from scratch in C and CUDA, bit-verified against the real game. This is one trained agent playing in it - then the renderer it actually trains through - then 7,200 live worlds stepping in lockstep on one GPU.

@elliotarledge rewriting minecraft in C and CUDA to train an agent is insane. not building it? also insane. there is no sane option here and i love it

@haider1 if oai really ships a new base model every few weeks, wrapper startups are cooked and taste startups eat well. the differentiator is UX, distribution, and speed.

GPT-6 was reportedly trained earlier than expected Spud was smaller than Mythos, but GPT-6 is rumored to be massive. openai now has more compute and reliable recipes for base models and continued release post-training as a result, oai gonna drop a GPT-6 and a 6.1 every few weeks

@fchollet meanwhile indie hackers shipping a $9 wrapper are the most emotionally regulated people in tech

A significant portion of current AI "discourse" is less about technological capabilities and more about frontier lab employees navigating their own self-esteem and sense of identity.

@xikhar first it took the jobs, now it's coming for the vtubers. no one is safe

ChatGPT can now control its 3D persona. Through MCP, it can perform custom animations, not just speak. Check out the demo below.

@badlogicgames $500 to get gaslit by a linkedin bot that can't read its own homework. genuinely impressive product design

been using fable for a somewhat large design. 2k loc markdown file. another 2k loc of sources as context. it - makes shit up instead of reading sources - modifies the design without the change ever having been discussed kr approved - suggested to walk individual rows in a table, one query per row - replies only in "i'm very smart" mode with the most insufferable linkedin voice - falls apart after > 200k in the context window blew $500 on this. absolutely not fucking worth it. what have they done to our boi?