<p class="f_center"><img src="http://dingyue.ws.126.net/2024/0305/cg00s9u6sh0066d200hs00kpg00g200io.gif"/><br/></p><p id="2HF9GGQF">刚刚,Anthropic发布了其新一代大语言模型Claude 3系列,包括Claude 3 Opus、Sonnet和Haiku三种规模,分别代表了超大杯、大杯和中杯。其中Sonnet版本在官网可以免费体验:https://claude.ai</p><p class="f_center"><img src="https://nimg.ws.126.net/?url=http%3A%2F%2Fdingyue.ws.126.net%2F2024%2F0305%2Fbcc83940j00s9u6si0020d200rm00m0g00g200cs.jpg&thumbnail=660x&quality=80&type=jpg" width="578" height="460" onload="this.removeAttribute('width'); this.removeAttribute('height'); this.removeAttribute('onload');" /><br/></p><p id="2HF9GGQH">而最强的Opus版本,则需要付费订阅,价格为20美元/月。</p><p id="2HF9GGQI">Anthropic声称,其中最强大的Claude 3 Opus模型在行业基准测试中超越了OpenAI的GPT-4和谷歌的Gemini Ultra,表现出了出色的知识理解和推理能力。</p><p class="f_center"><img src="https://nimg.ws.126.net/?url=http%3A%2F%2Fdingyue.ws.126.net%2F2024%2F0305%2F84174d68j00s9u6sk009od200pl00xpg00id00o6.jpg&thumbnail=660x&quality=80&type=jpg" width="661" height="870" onload="this.removeAttribute('width'); this.removeAttribute('height'); this.removeAttribute('onload');" /><br/></p><p id="2HF9GGQK">官方给出了一份42页的文档,感兴趣的可以去看一看:</p><p id="2HF9GGQL">https://www-cdn.anthropic.com/de8ba9b01c9ab7cbabf5c33b80b7bbc/Model_Card_Claude_3.pdf</p><p id="2HF9GGQM">抛开官方演示,Claude 3是首次提供多模态支持功能。用户可以上传照片、图表、文档等非结构化数据,由AI模型进行分析和回答。我们自然要来试一试。</p><p id="2HF9GGQN">比如给他一张图片,让他描述一下:<br/></p><p class="f_center"><img src="https://nimg.ws.126.net/?url=http%3A%2F%2Fdingyue.ws.126.net%2F2024%2F0305%2Fc0e2a099j00s9u6sl00d0d200u000kng00id00cm.jpg&thumbnail=660x&quality=80&type=jpg" width="661" height="454" onload="this.removeAttribute('width'); this.removeAttribute('height'); this.removeAttribute('onload');" /><br/></p><p id="2HF9GGQP">或是给他一道菜,让它给出这道菜的做法:<br/></p><p class="f_center"><img src="https://nimg.ws.126.net/?url=http%3A%2F%2Fdingyue.ws.126.net%2F2024%2F0305%2Fdba6922dj00s9u6sm001zd200u000rcg00g200em.jpg&thumbnail=660x&quality=80&type=jpg" width="578" height="526" onload="this.removeAttribute('width'); this.removeAttribute('height'); this.removeAttribute('onload');" /><br/></p><p id="2HF9GGQR">这个案例充分的证明了Claude 3并不理解中餐的魅力,这么一道西红柿炒鸡蛋愣是做成了米其林式的番茄鸡蛋汤,不仅用上了黄油橄榄油,还需要香草的点缀。<br/></p><p id="2HF9GGQS">或是给他一个手写的笔记,让他转录,似乎也还可以:<br/></p><p class="f_center"><img src="https://nimg.ws.126.net/?url=http%3A%2F%2Fdingyue.ws.126.net%2F2024%2F0305%2Faa1cd96cj00s9u6sn009kd200u000k5g00id00cb.jpg&thumbnail=660x&quality=80&type=jpg" width="661" height="443" onload="this.removeAttribute('width'); this.removeAttribute('height'); this.removeAttribute('onload');" /><br/></p><p id="2HF9GGQU">但是当笔记稍微潦草一点,就搞不定了:<br/></p><p class="f_center"><img src="https://nimg.ws.126.net/?url=http%3A%2F%2Fdingyue.ws.126.net%2F2024%2F0305%2F20b31c31j00s9u6sq00lwd200p000zgg00id00q1.jpg&thumbnail=660x&quality=80&type=jpg" width="661" height="937" onload="this.removeAttribute('width'); this.removeAttribute('height'); this.removeAttribute('onload');" /><br/></p><p id="2HF9GGR0">这次在多模态中也加入了代码生成能力,比如我给他一个页面,让他帮我生成实现这个页面的代码:<br/></p><p class="f_center"><img src="https://nimg.ws.126.net/?url=http%3A%2F%2Fdingyue.ws.126.net%2F2024%2F0305%2F4a38af96j00s9u6st00ajd200u000s6g00id00h8.jpg&thumbnail=660x&quality=80&type=jpg" width="661" height="620" onload="this.removeAttribute('width'); this.removeAttribute('height'); this.removeAttribute('onload');" /><br/></p><p id="2HF9GGR2">但是目前Claude 3 对于图片中复杂逻辑题的处理依然不理想,比如图片内容是一道物理题:</p><p class="f_center"><img src="https://nimg.ws.126.net/?url=http%3A%2F%2Fdingyue.ws.126.net%2F2024%2F0305%2Fj00s9u6sv00fed200u000lag00id00d0.jpg&thumbnail=660x&quality=80&type=jpg" width="661" height="468" onload="this.removeAttribute('width'); this.removeAttribute('height'); this.removeAttribute('onload');" /><br/></p><p id="2HF9GGR4">Claude 3 出现了幻觉,回答了一些并不存在的问题。</p><p id="2HF9GGR5">同样的问题GPT-4虽然也没能完全回答问题,但比Claude要强一些。</p><p class="f_center"><img src="https://nimg.ws.126.net/?url=http%3A%2F%2Fdingyue.ws.126.net%2F2024%2F0305%2F3edb4031j00s9u6sw005ad200u000p4g00id00fd.jpg&thumbnail=660x&quality=80&type=jpg" width="661" height="553" onload="this.removeAttribute('width'); this.removeAttribute('height'); this.removeAttribute('onload');" /><br/></p><p id="2HF9GGR7">虽然这次更新减少了拒绝回答的次数,但他的道德感依然要比GPT-4要强不少。<br/></p><p id="2HF9GGR8">比如让他做一张Web页面的UI代码,因为道德的原因被拒绝了。</p><p class="f_center"><img src="http://dingyue.ws.126.net/2024/0305/d53e3e98g00s9u6sy07ezd200dp007pg00id00ab.gif"/><br/></p><p class="f_center"><img src="https://nimg.ws.126.net/?url=http%3A%2F%2Fdingyue.ws.126.net%2F2024%2F0305%2F6e215c2aj00s9u6t000bed200u0009xg00id0062.jpg&thumbnail=660x&quality=80&type=jpg" width="661" height="218" onload="this.removeAttribute('width'); this.removeAttribute('height'); this.removeAttribute('onload');" /><br/></p><p id="2HF9GGRB">案例来自@RubenHssd</p><p id="2HF9GGRC">除了多模态能力,这次更新后的长文本能力。目前支持200K Token上下文,未来可能会支持到1M的规模。</p><p id="2HF9GGRD">通过QuALITY测试,Claude 3 Opus在1-shot(一次提示)设置下达到了90.5%的准确率,在0-shot(无提示)设置下达到了89.2%的准确率。</p><p class="f_center"><img src="https://nimg.ws.126.net/?url=http%3A%2F%2Fdingyue.ws.126.net%2F2024%2F0305%2Fab1035dfj00s9u6t10015d200ry0064g00id0040.jpg&thumbnail=660x&quality=80&type=jpg" width="661" height="144" onload="this.removeAttribute('width'); this.removeAttribute('height'); this.removeAttribute('onload');" /><br/></p><p id="2HF9GGRF">据说当时GPT-3发布就是OpenAI听说Anthropic要发布Claude;在Claude2发布的时候,OpenAI用被称作GPT-4.5的Code Interpreter作为反击;在Anthropic宣布获得亚马逊40亿美元融资的时候,OpenAI为ChatGPT开放了语音和图像能力。<br/></p><p class="f_center"><img src="https://nimg.ws.126.net/?url=http%3A%2F%2Fdingyue.ws.126.net%2F2024%2F0305%2Fef002421j00s9u6t1001ud200lf00p0g00id00lf.jpg&thumbnail=660x&quality=80&type=jpg" width="661" height="771" onload="this.removeAttribute('width'); this.removeAttribute('height'); this.removeAttribute('onload');" /><br/></p><p id="2HF9GGRH">可以说Claude的进展就是OpenAI要有新动作的预告片。<br/></p><p id="2HF9GGRI">在临近截稿的时候,OpenA宣布ChatGPT增加了一个新功能——阅读回答。</p><p class="f_center"><img src="https://nimg.ws.126.net/?url=http%3A%2F%2Fdingyue.ws.126.net%2F2024%2F0305%2Fb20275dcj00s9u6t20064d200p100rjg00id00k7.jpg&thumbnail=660x&quality=80&type=jpg" width="661" height="727" onload="this.removeAttribute('width'); this.removeAttribute('height'); this.removeAttribute('onload');" /><br/></p><p id="2HF9GGRK">网友对这个动作十分不满意,包括我:<br/></p><p class="f_center"><img src="https://nimg.ws.126.net/?url=http%3A%2F%2Fdingyue.ws.126.net%2F2024%2F0305%2F4075ed13j00s9u6t3002zd200ou00hfg00id00cv.jpg&thumbnail=660x&quality=80&type=jpg" width="661" height="463" onload="this.removeAttribute('width'); this.removeAttribute('height'); this.removeAttribute('onload');" /><br/></p><p id="2HF9GGRM">甚至连嘲讽的梗图都出来了:<br/></p><p class="f_center"><img src="https://nimg.ws.126.net/?url=http%3A%2F%2Fdingyue.ws.126.net%2F2024%2F0305%2Ffd7d0dfcj00s9u6t500eud200o300gmg00id00cn.jpg&thumbnail=660x&quality=80&type=jpg" width="661" height="455" onload="this.removeAttribute('width'); this.removeAttribute('height'); this.removeAttribute('onload');" /><br/></p><p id="2HF9GGRO">这样是不是可以理解为OpenAI地主家也没有存粮来狙击Anthropic了。<br/></p>
GPT plus 代充 只需 145
版权声明:本文内容由互联网用户自发贡献,该文观点仅代表作者本人。本站仅提供信息存储空间服务,不拥有所有权,不承担相关法律责任。如发现本站有涉嫌侵权/违法违规的内容,请联系我们,一经查实,本站将立刻删除。
如需转载请保留出处:https://51itzy.com/kjqy/210344.html