Sebastien Ego
68 posts
Sebastien Ego
@SebastienEgo
AI, SaaS & startups. French solo founder.
Katılım Haziran 2026
72 Takip Edilen13 Takipçiler
Rejourney runs its issue detection on Gemini for cost, GPT-5.5 if you self-host. The LLM is now a swappable backend, not the product. The second-order effect: the moat is the cohort heuristic and the session data pipeline, not model choice.
English
@pato__project Combo orchestrateur / exécutant : Fable + Sonnet ou Fable + GPT 5.6 Sol
Français
@phibrowser @OfficialLoganK co-authorship still assumes symmetric context though. the human sees the whole diff, does the agent see the human's uncommitted edits in real time or only at handoff? that asymmetry decides who's actually driving.
English
@SebastienEgo @OfficialLoganK 'primary author' is the load-bearing assumption, and the wrong axis. the shift isn't agent-first vs human-first, it's co-authorship: control passing back and forth in one surface. the environment is for two authors, not a handoff of which is primary.
English
The Agentic Coding Environment (ACE) is the natural successor to the IDE (Integrated Development Environment)
English
@pcshipp tried the same on a Mac Studio with Qwen 3.6 and Gemma 4, and it falls apart fast on anything harder than basic tasks. curious to see how it holds up for you though.
English
@Haleeeemahh 30-46% on OpenRouter means companies aren't picking a model, they're building a router that picks per-task. That's the actual moat question: whose orchestration layer decides when "good enough" is really good enough
English
Sam Altman’s response to questions about chinese Ai models
If you want the best results at higher stick to frontier models
If you want good results at cheap cost/free stick to chinese open weights models
:) 🤝
Haleemah@Haleeeemahh
chinese ai models vs frontier models companies are starting to route simpler tasks away from claude and toward cheaper chinese models to save money chinese models like deepseek v4, glm 5.2, and qwen reportedly now handle 30-46% of tokens used by us companies on openrouter although some are still a few months behind the best frontier models in performance but for tasks that don’t need the absolute best model, chinese models are winning on cost everyone wants fable 5 right now… but when it switches to paid credits in a few days, will you still use it? or will you switch to a cheaper model?
English
@emollick the frontier-less vendors also tend to have the most detailed slides on why benchmarks don't matter.
English
Lots of the AI & corporate strategy talk on X is not that complicated: vendors without frontier models are always explaining why you shouldn't trust the frontier model companies. Vendors w/frontier models are selling solutions based around frontier models. (Either could be right)
English
Quiet winner of OpenAI's week: Anthropic. Zero hardware bets, zero Apple lawsuits, just model access and identity checks. Quiet loser: any solo builder who shipped on Atlas or bet a product roadmap on OpenAI hardware timing.
ai-supremacy.com/p/openai-just-…
English
English
@garrytan and @BrianRoemmele have both been addressing this in a different manner.
Owning your data, not depending on a single-source or frontier model, using open source local AI, creating a sandbox.
So much to learn and do.
Satya Nadella@satyanadella
English
@AndrewCurran_ 200 signatures and the word "prepare" doing all the work, no timeline attached. Compare it to climate statements from a decade ago, same structure, same vagueness on who moves first, governments or labs.
English
More than 200 researchers and economists including Jack Clark, Jeff Dean, Noam Brown, Tyler Cowen, Sholto Douglas, John Schulman, Yoshua Bengio, Eric Schmidt, and Dean Ball have signed a statement urging governments and institutions to act now to prepare for AI's economic impact.
English
@BrianRoemmele 82.7 SWE-bench with free API access day one, that's the part I'd want the mechanism on. Who's footing inference cost at that price, Singapore state backing or just burning VC runway to get distribution before pricing kicks in?
English
Boom!
I am testing soon to be open source and currently FREE Singapore-based Agnes 2.5 Pro and it delivers frontier-level coding performance, scoring 82.7 on SWE-bench Verified, 78.7 multilingual, and strong results on SWE Atlas and Terminal-Bench!
The model outperforms GLM 5.2 and DeepSeek V4 Pro across multiple benchmarks, with charts showing Agnes variants leading against Claude Opus 4.8 and other leaders in internal evaluations!
Release marks a shift in global AI, expanding frontier model development beyond the US-China dynamic as more countries enter the race, with free API access available immediately.
English
@copilot_shogo free stress-test, paid for in extended patience from everyone hitting rate limits anyway.
English
Interesting move from Anthropic.
They've extended Fable 5 access again, along with the 50% higher Claude Code rate limits.
My guess is that their infrastructure is handling demand better than expected. Every extra week gives them more real-world usage data, helps validate GPU capacity, and lets more developers build Fable 5 into their workflow before usage-based pricing kicks in.
A smart way to stress-test a frontier model at scale while increasing long-term adoption.
Claude@claudeai
We're extending Claude Fable 5 access on all paid plans, as well as keeping Claude Code’s weekly rate limits 50% higher, through July 19.
English
distillation bans and identity checks are the same move: gate the inference layer, not the weights. open models don't have this problem, that's the actual moat shift builders should track.
x.com/maxforai/statu…
Max For AI@MaxForAI
据《金融时报》报道称，OpenAI宣布已暂停部分与阿里巴巴有关联账户的访问权限，原因是发现其涉嫌利用模型输出进行“蒸馏”。 OpenAI 表示，这些账户存在利用模型输出训练其他模型等违规行为，并表示正在持续加强对相关风险的识别和防范措施。 Google 表示其 AI 模型同样提供给阿里巴巴、百度、腾讯位于新加坡的子公司使用。 由于这些海外子公司本身未被列入禁止交易的美国出口管制清单，Google 提供服务并未违反现行规定。
English
Everyone's reading the CDC proof as "AI is now doing novel math." That's not the story. The story is a PDF from a closed lab is the only artifact anyone gets to inspect. No open weights model shipped this, so nobody outside OpenAI can rerun the reasoning trace and check if it's real proof or curve-fit theorem-shaped text. Watch whether an independent group reproduces it before you update your priors on model capability.
cdn.openai.com/pdf/04d1d1e4-b…
English
@elonmusk /loop 5m /pr-babysit check until the sun burns out, the real end-to-end workflow.
English
Europe gets frontier models last, again. Pattern's been the same for a while: US ships, rest of world waits on "compliance review."
Meanwhile the release cadence keeps compressing. Distribution lag is starting to matter as much as the model itself.
aiweekly.co/alerts/spacexa…
English
@WriterLamin every model launch season some feature quietly gets sacrificed to the update gods, this time it was Fable 5's turn.
English
My Claude opus 4.8, ever since Grok 4.5 and ChatGPT 5.6 announced new updates... model OVERLOAD!
My Fable 5 is gone
Dear Anthropic, you are giving me a BAD DEAL. WHAT'S UP!?
English
@imhaoyi GLM-5.2 comparison is the interesting part, not the benchmark score. in actual multi-step agent runs does it feel closer to Opus 4.8 or does it drift more once context gets long?
English
SpaceXAI dropped Grok 4.5 and it's actually good enough to pay attention to now.
Coding and engineering: hangs with GPT 5.5 and Claude Opus 4.8. SWE Bench Pro hit 64.7%, first time beating GPT 5.5 on this one. In practice I'd put it around GLM-5.2 level.
$2/M input, $6/M output. Cheap relative to the rest of the frontier.
English